Exploring Large Language Models' Responses to Moral Reasoning Dilemmas

Exploring Large Language Models' Responses to Moral Reasoning Dilemmas | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Exploring Large Language Models' Responses to Moral Reasoning Dilemmas Davin Nabizadeh, David Walker, Hyemin Han, Emily Laird This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6823916/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract This study investigates how various large language models (LLMs) generate responses to moral reasoning dilemmas. It specifically examines LLM-generated responses using the Defining Issues Test (DIT-2) and the Intermediate Concepts Measure (ICM) for Educational Leaders. Using a neo-Kohlbergian approach to moral reasoning, the study evaluates responses from multiple LLM platforms: ChatGPT-3.5, ChatGPT-4, ChatGPT-4O, Grok Premium Plus, Claude 3.5 Sonnet, Gemini, and Gemini Advanced. For DIT-2, Claude learns to prioritize the highest post-conventional moral reasoning score and N2 score (P-score 72, N2 score 71.10), followed by Gemini Advanced (P-score 64, N2 score 60.31) and Gemini (P-score 58, N2 score 52.11). Other LLMs performed as follows: Grok (P-score 48, N2 score 47.98), ChatGPT-4O (P-score 44, N2 score 55.07), ChatGPT-4 (P-score 44, N2 score 46.53), and ChatGPT-3.5 (P-score 18, N2 score 36.20). For the ICM Educational Leaders version, Gemini Advanced had the highest total ICM score of 0.90, followed by Claude 3.5 Sonnet and Gemini (both 0.86), ChatGPT-4O and ChatGPT-4 (both 0.78), Grok (0.61), and ChatGPT-3.5 (0.32). The findings indicate that some LLMs can generate responses consistent with sophisticated moral reasoning patterns, producing scores comparable to or exceeding graduate-level human participants (whose P-scores typically range from 38.5 to 42.3) and provide a methodological framework consisting of standardized assessment protocols and comparative analysis techniques for larger-scale research to improve our understanding of AI's potential in moral reasoning. Large Language Models (LLMs) Moral Reasoning Defining Issues Test (DIT-2) Intermediate Concept Measure (ICM) Introduction Artificial intelligence (AI) originated in 1956 when John McCarthy organized a two-month workshop at Dartmouth College and for the first time introduced the term "artificial intelligence" (McCarthy et al., 2006 ). AI simulates human intelligence using algorithms, data, and computing power (Haenlein & Kaplan, 2019 ). Since then, AI has developed in various fields, for instance in education, with applications ranging from intelligent tutoring systems to personalized learning and administrative support in higher education (Roberts & Park, 1983 ; Zawacki-Richter et al., 2019 ). The evolution of AI has progressed from simple rule-based systems including statistical language models to neural language models and then to advanced deep learning models (Zhao et al., 2023 ). As an example, in late 2022, OpenAI introduced ChatGPT, a powerful LLM that revolutionized AI applications across industries, excelling in text generation, coding, multilingual processing, and advanced reasoning (Alawida et al., 2023 ; Bansal et al., 2024 ; de Winter et al., 2024 ; Wu et al., 2024 ; Zhang et al., 2024 ). LLMs like ChatGPT are built on advanced neural network architectures that use deep learning and natural language processing (NLP). They employ advanced NLP, supervised learning, and reinforcement learning techniques to understand and produce text that simulates human conversation (Roumeliotis & Tselikas, 2023 ). These models have been trained on massive amounts of data, allowing them to respond to a variety of questions in ways that are comparable to humans. These models' ongoing improvements, particularly those developed after late 2022, reflect advances in AI research across multiple dimensions: architectural enhancements (improved transformer designs, attention mechanisms), training methodology refinements (reinforcement learning from human feedback, constitutional AI approaches), increased parameter counts and training data scale, and improved performance metrics on reasoning benchmarks and AI systems' growing ability to understand and generate complex language (Chu et al., 2024 ; Roumeliotis & Tselikas, 2023 ; Zubiaga, 2024 ). The rapid advancement of AI technologies has increased interest in comparing AI systems to humans in various domains, including moral reasoning and how human actors perceive and construe the moral persona of AI teammates (Han, 2023 ; Sengupta et al., 2025 ; Tanmay et al., 2023 ). While the benefits of LLMs in natural language processing and automation are well known, their potential to influence and challenge complex fields such as morality and moral reasoning is opening up new avenues for understanding human ethics, human-AI teaming, and improving decision-making frameworks (Jorgenson et al., 2025 ; Ramezani & Xu, 2023 ; Sengupta et al., 2025 ; Simmons, 2022 ; Slavkovik, 2022 ). Even though LLMs do not represent moral concepts in the same way as humans do, their training on large amounts of textual data provides them with a statistical understanding of moral concepts such as fairness and justice that are common in human societies (Pock et al., 2023 ). AI may reflect moral values, which challenges the idea that all technologies are inherently value-neutral (Swoboda & Lauwaert, 2025 ). Modern LLMs incorporate sophisticated safety features and alignment mechanisms, including reinforcement learning from human feedback (RLHF) and constitutional AI approaches, which are designed to steer outputs toward ethical behavior and away from harmful content. Unaligned LLMs (these models not programmed with specific human moral values, reflecting diverse societal biases) understand moral concepts conceptually because of their exposure to the social totality, but their moral reasoning is still limited and lacks the depth of human understanding (Pock et al., 2023 ). The widespread use of LLMs in various industries highlights the importance of understanding their moral reasoning capacities. Since these models are increasingly integrated into decision-making processes in healthcare (Armitage, 2025 ; Hager et al., 2024 ), education (Zhui et al., 2024 ), legal systems (Almeida et al., 2024 ; Marcos, 2024 ), art (Ivanova, 2025 ), and self-driving cars (Ahmad & Takemoto, 2024 ; Takemoto, 2024 ; Xu et al., 2025 ), their ability to navigate ethical complexities is critical. For example, AI systems in healthcare must navigate complex ethical considerations involving patient privacy, diagnostic accuracy, and resource allocation, where modern privacy-preserving techniques such as federated learning and differential privacy demonstrate that these concerns can be addressed without simple trade-offs (Armitage, 2025 ; Hager et al., 2024 ), and self-driving cars must make quick decisions that could cost lives (Ahmad & Takemoto, 2024 ; Awad et al., 2018 ; Takemoto, 2024 ; Xu et al., 2025 ). These real-world applications highlight the importance of examining LLMs' functional moral reasoning capabilities and ensuring that they are consistent with human values. Furthermore, as educational institutions increasingly adopt AI for personalized learning and administrative tasks (Gan et al., 2023 ; Xing et al., 2025 ; Yan et al., 2024 ), understanding how LLMs generate responses to moral dilemmas becomes critical for developing responsible AI systems that promote rather than interfere with ethical development. By exploring how different LLMs respond to established moral reasoning tests, this study provides insight into the capabilities and limitations of current AI systems, as well as providing foundations for future research and development in this critical area. As a result, this research aims to determine how LLMs generate responses consistent with moral reasoning patterns across various versions of LLMs to identify differences, similarities, and potential insights into how AI simulates human moral reasoning. Finally, to provide a more comprehensive understanding of LLMs' moral reasoning response capabilities, this study examines the correlation between abstract moral reasoning (measured by DIT-2) and domain-specific professional moral reasoning (measured by ICM). This correlation analysis is essential because it reveals whether LLMs generate coherent moral reasoning response patterns that function consistently across different contexts and levels of abstraction or whether their moral reasoning is inconsistent and context dependent. For example, previous studies with human participants showed significant correlations between abstract moral reasoning and its application in professional contexts (Thoma et al., 2008 ). This indicates that established moral reasoning systems should be consistent across measures. As a result, by exploring these correlations in LLMs, researchers can learn whether improvements in one aspect of moral reasoning response generation correspond to improvements in others, which has important implications for the overall development of ethical AI systems. The Importance of AI in Moral Reasoning LLMs can generate responses consistent with moral reasoning patterns using top-down approaches based on normative and descriptive ethical theories, such as, normative ethical theories (deontology, utilitarianism) and descriptive frameworks (Theory of Dyadic Morality). Top-down approaches start with predefined ethical principles, rules, or theories, guiding decision-making and moral judgments, while bottom-up systems develop ethical understanding through experience, or crowd-sourced inputs, observing patterns in data. These approaches provide better explainability and structure than traditional bottom-up approaches, which learn ethics from data without explicit moral guidelines (Zhou et al., 2023 ). In addition, recent research applies virtue ethics to artificial agents. This mostly bottom-up approach emphasizes moral character development, practical wisdom, and experience-based learning, and it provides another avenue for artificial agents to adapt to complex moral dilemmas (Stenseke, 2023 , 2024 ). The application of the aforementioned ethical frameworks in AI systems creates major challenges. Utilitarian theories may encounter difficulties in quantifying and evaluating benefit across diverse outcomes, whereas deontological frameworks may become excessively inflexible for complex real-world situations. Conversely, virtue ethics offers a compelling alternative by prioritizing the cultivation of moral character and practical wisdom; however, the translation of these principles into computer models poses significant challenges due to the complexity of different situations and the context-sensitivity required for phronesis (practical wisdom) to function properly (Stenseke, 2023 , 2024 ). As AI systems become more effectively incorporated into various aspects of daily life, the significance of AI systems that can represent and respond to human values and engage in moral reasoning and judgment increases (Peterson & Gärdenfors, 2024 ; Sullivan & Fosso Wamba, 2022 ). The potential challenges of using AI to make ethical decisions have been the subject of recent research (Slavkovik, 2022 ). For example, AI in autonomous vehicles needs to be able to make nuanced moral decisions in life-or-death circumstances. These decisions include identifying least harmful options in unavoidable accident scenarios such as whether to prioritize passenger over pedestrian safety (Awad et al., 2018 ). The capacity of AI to engage in moral reasoning in these circumstances is both a technical challenge and an ethical responsibility. Moreover, the cultural specificity of moral norms adds another layer of complexity to AI moral reasoning. What is considered ethical can vary significantly across different cultures and contexts, which creates concern about whose values should be encoded in AI systems. As a result, this challenge is particularly relevant for LLMs trained on diverse datasets that may contain conflicting moral perspectives. Research has shown that English pre-trained Language models (EPLMs) generate a certain level of moral knowledge but are heavily biased toward western values, making it difficult for them to generalize to diverse cultural contexts (Jinnai, 2024 ; Ramezani & Xu, 2023 ). Evaluating LLMs' moral reasoning response capabilities creates several methodological challenges. Unlike humans, LLMs do not have personal experiences or emotions to guide their moral reasoning. On the contrary, LLMs rely only on the patterns identified in their training data, which may include biased or contradictory moral perspectives. Furthermore, LLMs may generate responses that appear morally sophisticated but lack the fundamental understanding of philosophical foundations, potential implications, and contextual nuances that distinguishes human moral reasoning (Bonagiri et al., 2024 ). Another limitation is the possibility of "moral mimicry," in which LLMs generate moral rationalizations that mimic human reasoning without engaging with the complex moral implications involved in a situation. For example, research found that LLMs can produce moral biases related to political identities in morally loaded responses to political prompts, which indicates that these models simulate patterns from their training data rather than engaging in genuine moral reasoning (Simmons, 2022 ). Despite these limitations, exploring LLMs' responses to standardized moral reasoning tests provides useful information about their capabilities and potential applications. In addition, through comparing different models across various measures and examining the patterns in their answers, researchers can obtain a deeper understanding of how these systems process and generate moral judgments, find areas for improvement, and develop more ethically aligned AI systems. Previous research on human participants has demonstrated that sophisticated moral reasoning is necessary for societal functioning and personal development. Higher levels of moral reasoning have been linked to prosocial behavior, ethical decision-making in professional settings, and the ability to solve complex moral dilemmas (Kohlberg & Power, 1981 ; Rest et al., 2000 ). In democratic societies, the capacity to engage in post-conventional moral reasoning which prioritizes universal ethical principles over mere rule-following enables critical evaluation of social norms and institutions, which results in contributing to social progress and justice. For instance, individuals with advanced moral reasoning abilities can make decisions that balance competing ethical principles while also considering multiple stakeholders, which is increasingly important in today's complex, pluralistic world. Creating AI systems with sophisticated moral reasoning response capabilities is important as they become more integrated into decision-making processes across multiple domains. As a result, by exploring how various LLMs respond to moral reasoning evaluations, researchers can gain a better understanding of their capabilities and limitations, identify areas for improvement, and create more ethically aligned AI systems capable of navigating complex moral dilemmas in ways that are consistent with human values. A recent study used the Behavioral Defining Issues Test to assess ChatGPT's moral reasoning capacity. The findings revealed that ChatGPT provided moral reasoning responses equal to human participants, scoring 45.83 in post-conventional reasoning (slightly lower than the median score of 50.00 achieved by human undergraduate students, with a trivial effect size difference). These findings indicate that LLMs such as ChatGPT have the potential to generate responses that functionally resemble features of human moral reasoning (Han, 2023 ). In a comparative study of moral decision-making, narratives generated by human participants and ChatGPT-3 were assessed using metrics such as causality, explicability, and overall satisfaction. The study showed no significant difference in the quality of explanations between human and AI-generated responses, which indicates that ChatGPT-3 can generate moral justifications comparable to those of humans (Rehman et al., 2025 ). Another group of researchers used the Defining Issues Test (DIT-1) to assess the moral reasoning abilities of LLMs, such as GPT-3, GPT-3.5, GPT-4, ChatGPT (July 2023), Llama2-Chat, and PaLM-2. The results varied among models; for example, GPT-3 performed closest to a random baseline, indicating inadequate moral reasoning abilities. However, GPT-4 performed well, with a P-score of 55.68, indicating a post-conventional level of moral reasoning equivalent to that of graduate students (advanced level explained further below). In addition, ChatGPT, Llama2-Chat, and PaLM-2 showed lower reasoning abilities than young adults or college students. The study emphasized GPT-4's significant improvements over prior versions but also found inconsistency across various scenarios. Researchers proposed that the models' reasoning abilities were the result of extensive training data and improved Reinforcement Learning from Human Feedback (RLHF) (Tanmay et al., 2023 ). Studies on LLMs trained with RLHF demonstrate that models can avoid harmful outputs such as bias and discrimination when given clear ethical instructions. This ability to generate responses that align with ethical guidance through natural language instructions is a positive sign for the future of LLMs in moral reasoning tasks (Askell et al., 2021 ; Ganguli et al., 2023 ; Ganguli et al., 2022 ; Gehman et al., 2020 ). In essence, LLMs can be trained to produce responses that follow moral reasoning protocols by using their capacity to internalize ethical instructions learned through human feedback (Ganguli et al., 2023 ). This area of study continues to grow, and there are challenges such as enabling AI systems to perform moral reasoning, bias, and transparency (Slavkovik, 2022 ). This concern highlights the need for additional development of LLMs to better align with human moral frameworks (Pock et al., 2023 ; Slavkovik, 2022 ). This study addresses this gap by exploring the generative responses of various LLMs to well-known moral reasoning tools such as the Defining Issues Test (DIT) and Intermediate Concept Measure (ICM) (Kerr, 2018 ; Thoma & Dong, 2014 ). The present article explores the degree to which LLMs generate responses that functionally resemble human moral reasoning and provides an extensive evaluation of the compatibility of various LLM versions with human moral reasoning frameworks. Neo-Kohlbergian Theory of Human Moral Functioning The study reported in this article explores LLMs' responses to moral reasoning dilemmas through the lens of neo-Kohlbergian theory including the Four-Component Model (FCM) (Rest, Thoma, et al., 1999). FCM explains morality through four psychological processes (Narvaez & Rest, 1995 ; Rest, 1994a ). First, Moral sensitivity involves detecting potential harms to others, being concerned about others' welfare, and recognizing ethical issues. Second, moral judgment determines right and wrong actions and provides rationales for behavioral decisions. Moral reasoning involves weighing up moral features of a situation toward deciding how to respond. Third, moral motivation prioritizes moral values over personal gains or social approval as reasons for one’s actions, and finally, moral character involves acting on moral values and judgments. Each component of the FCM plays a distinct but interconnected role in moral functioning (Rest, 1994a , 1994b ). The focus of the reported study is on the moral reasoning component where most neo-Kohlbergian expertise can be found. Moral reasoning is the cognitive process that happens when individuals make decisions about what is morally right or wrong in a particular situation (Kohlberg & Power, 1981 ; Richardson, 2003 ); for example, a father contemplates stealing food for his starving family from the warehouse of a rich man hoarding food (Thoma & Dong, 2014 ). Within the component of moral judgment, there are three levels: codes of conduct, intermediate concepts, and bedrock schemas. Codes of conduct are specific rules requiring minimal interpretation; intermediate concepts involve virtue-like concepts involving contextualized understanding, such as concepts of informed consent or virtues such as honesty and courage. Intermediate concepts may be evaluated by Intermediate Concept Measures (ICMs) that need to be developed uniquely for a context such as a profession (teaching) or stage of development (adolescent). ICMs are, therefore, uniquely focused on practical contextualized moral reasoning (Thoma et al., 2013 ). Bedrock moral schemas are considered global schemas involving a person’s general moral orientation to the world. Developed from Kohlberg’s original stage-based theory moral schemas are evaluated by the Defining Issues Test (DIT). The DIT measures moral reasoning and activates three moral schemas: personal interests, maintaining norms, and post-conventional (Rest, Narvaez, Bebeau, et al., 1999; Rest et al., 2000 ). The neo-Kohlbergian approach underpinning this perspective, places particular emphasis on progression to a post-conventional level of moral reasoning in adolescents and adults as the highest level of moral reasoning. Dropping Kohlberg’s original stage-based approach for moral schemas allows for fluidity between ‘levels’ such that individuals may use different schemas depending on the situation, which leads to a more profound understanding of moral development (Thoma & Dong, 2014 ). At the lowest level of moral reasoning, the personal interest schema involves viewing oneself and others as motivated by individual benefits and includes efforts to maintain approval from others. In this way, self-interest and social approval trump broader ethical principles. The maintaining norms schema emphasizes the importance of established rules, roles, and formal organizational structures. More advanced than personal interests but still limited, it focuses on maintaining the current social order and adhering to established procedures and authorities. The post-conventional schema as the most advanced form of moral reasoning involves nuanced and sophisticated moral reasoning capable of exceeding local norms if they curtail ethical goods. The DIT P-score represents the likelihood of utilizing the postconventional schema, focusing on universal ideals and full reciprocity of social norms, which emphasizes principles that can be justified in a society committed to moral values. Overall, post conventional reasoning includes a critical examination of societal norms as well as a commitment to universal principles of justice, rights, and welfare (Rest, Narvaez, Bebeau, et al., 1999; Thoma, 2014 ). Neo-Kohlbergian theory, involving bedrock moral schemas and intermediate concepts, provides a framework for assessing and potentially improving LLM moral reasoning abilities, which is useful for AI development. In an ideal world, advanced AI systems would be able to engage in post-conventional reasoning, balancing universal ethical principles with contextual awareness. However, achieving this level of moral reasoning in AI systems presents significant challenges, particularly given the complexity and sometimes contradictory nature of human ethical reasoning as well as cultural differences. Researchers can gain a better understanding of specific AI’s current capabilities and identify areas for improvement in future AI development by examining how different LLMs perform on measures such as DIT-2 (as the latest version of DIT) and ICM. In applying neo-Kohlbergian theory to AI systems, some challenges emerge. For example, current LLMs can simulate aspects of moral judgment using pattern recognition in training data, but they lack the real-life understanding that drives human moral sensitivity and the internal motivational structures that drive moral action. LLMs can recognize morally relevant situations and respond in line with moral principles, but they do not "care" about moral outcomes in the same way humans do. Despite these limitations, studying LLMs' responses to moral reasoning tasks reveals important information about their ability to simulate moral judgment processes. Method The study used multiple LLM platforms alongside two moral reasoning measures (DIT-2 and ICM for Educational Leaders). In this section, we introduce the LLM platforms used in this study, describe the moral reasoning assessment tools, and explain the data collection and scoring procedures. LLMs Platforms The LLM models included in the research are: ChatGPT-3.5, ChatGPT-4, ChatGPT-4O, Grok Premium Plus, Claude 3.5 Sonnet, Gemini, and Gemini Advanced. These LLMs were chosen based on their popularity and recognition as of summer 2024, ensuring the study examined the most widely used models. The diversity of these models, with their unique architectures, training techniques, and design principles, provides a comprehensive sample for assessing moral reasoning abilities. As a result, by comparing their standardized moral reasoning test results, we can identify similarities and differences between model design decisions and moral reasoning abilities. A summary of each LLM is provided below. ChatGPT ChatGPT has evolved significantly since its initial release. For example, ChatGPT-3.5, which is based on the GPT-3.5 architecture, demonstrated significant improvements in natural language processing capabilities, but it lacked complex reasoning and consistency. ChatGPT-4, on the other hand, demonstrated significant improvements in reasoning, contextual understanding, and hallucination reduction. ChatGPT-4 also had improved multimodal capabilities, allowing it to process text and images with greater efficiency and lower latency (Alawida et al., 2023 ; Zhang et al., 2024 ). Claude 3.5 Sonnet Anthropic's Claude 3.5 Sonnet emphasizes helpfulness, harmlessness, and honesty. It employs Anthropic's constitutional AI approach, which seeks to align AI systems with human values while reducing harmful outputs. Claude 3.5 Sonnet performs well in tasks that require nuanced understanding, contextual reasoning, and careful handling of sensitive topics. Its training methodology emphasizes both capability and safety, with the goal of developing a balanced system capable of dealing with complex questions while adhering to ethical boundaries (Bae et al., 2024 ). Gemini DeepMind developed Google's Gemini models, which represent a significant advancement in multimodal AI capabilities. Gemini uses a unified architecture to process and reason about text, images, audio, and video. The standard Gemini model provides balanced performance across multiple tasks, whereas 'Gemini Advanced' improves reasoning abilities, particularly for complex analytical and creative tasks. Both models use sophisticated training methods that combine multimodal comprehension and advanced reasoning abilities (Team et al., 2023 ). Grok Premium Plus Grok, created by xAI under Elon Musk's direction, focuses on real-time information integration and a unique approach to engagement that includes humor elements. Grok Premium Plus, the enhanced version used in this study, has improved reasoning capabilities and access to current information. The model seeks to distinguish itself by adopting a more conversational and less restrictive interaction style while maintaining reasoning abilities comparable to other leading LLMs (Wangsa et al., 2024 ). Moral Reasoning Tools Defining Issues Test (DIT2) The Defining Issues Test (DIT) is a widely used assessment tool for measuring moral reasoning. DIT-2 is an updated version of the original DIT, incorporating more contemporary moral dilemmas and improved scoring algorithms. It presents participants with five moral dilemmas and asks them to rate and rank different responses based on their importance. In fact, the DIT test generates several indices, such as personal interest schema, maintaining norms schema and postconventional schema. The DIT has been extensively validated and is supported by many studies, which shows its strong reliability and validity (Gungordu et al., 2024 ; Thoma & Dong, 2014 ). DIT-2 includes several indices that help measure other aspects of moral reasoning. These indices are: The Type Indicator (TYPENEW) : the dominant moral schema is represented by the Type Indicator, which ranges from 1 to 7. Type 7 indicates the predominant postconventional schema that is consolidated. In contrast, Type 4 represents predominant maintaining norms schema that is consolidated. In addition, Types 1, 4, and 7 are consolidated profiles, while types 2, 3, 5, and 6 are transitional profiles. Utilizer Score (U) : The Utilizer Score (-1 to 1) measures consistency between moral reasoning and action choices. Scores closer to 1 indicate greater agreement between item endorsement and action choice, whereas scores closer to -1 indicate less agreement. Humanitarian Liberalism (HUMLIB) : This score (0–5) is based on the number of humanitarian liberal action choices made. Higher scores indicate stronger alignment with positions endorsed by experts in political science and philosophy. Cannot Decide Choices (NUMCD) : This indicates how many times "can't decide" was chosen across the dilemmas (1–5). Indecision can be linked to developmental stages, with transitional phases frequently showing increased indecision due to multiple conflicting interpretations. Religious Orthodoxy (CANCER10) : This is a proxy measure (1–9) based on item 10 in story 4 (the cancer story). Lower scores indicate less religious orthodoxy. It is calculated as the sum of the rating and weighted rank given to an item that evokes the notion that only God should determine whether someone lives or dies. Consolidation/Transition (CONSTRAN) : A score of 2 indicates a consolidated profile, meaning there is clear evidence of schema preference. A score of 1 would indicate a transitional profile, reflecting developmental disequilibrium with less discrimination among schema-typed items (Bebeau & Thoma, 2003 ). The four primary indices which represent the likelihood of utilizing each schema are the most important indices for the DIT2. As mentioned earlier, the Personal Interest Schema score (0-100) is the lowest level of moral reasoning. The Maintaining Norms schema score (0–98) is an improvement on that, whereas the post-conventional moral schema score (0–95) shows sophisticated moral reasoning. The P-Score is an overall index of schema consideration, and is the traditional score used in most DIT research. Finally, the N2 score (0–95) is a relatively new indicator that assesses both the prioritization of post-conventional items and the de-emphasis of personal interest items. N2 score provides a more nuanced understanding of moral development than just the P score (Bebeau & Thoma, 2003 ). There are some additional Indices. The Antisocial Score (ASCORE) : it ranges from 0 to 16, with higher scores indicating stronger antisocial/anti-establishment attitudes. These considerations assume understanding of stage 4, but criticize existing authorities for hypocrisy (also referred to as stage 4½). Meaningless Items (MSCORE) : Scales from 0 to 16, indicating endorsement of items with complex but meaningless wording. Higher scores may indicate attempts to fake higher scores and/or a lack of understanding. Participants who receive scores greater than 10 will be purged. New Checks (TOTCC) : A reliability check that combines multiple indicators to detect invalid responses. Participants who receive scores greater than 200 will be purged. It checks for rate-and-rank consistency, missing data, endorsement of meaningless items, and non-discrimination (giving the same rating to too many items). Purged Participants (PURGED) : A binary indicator (0 = Included, 1 = Purged) that indicates whether a respondent's data was retained or excluded from analysis due to reliability checks (Bebeau & Thoma, 2003 ). The DIT-2 comprises five moral dilemmas: (1) Famine : a father contemplates stealing food for his starving family from the warehouse of a rich man hoarding food; (2) Reporter : a newspaper reporter must decide whether to report a damaging story about a political candidate; (3) School Board : school board chair must decide whether to hold a contentious and dangerous open meeting; (4) Cancer : a doctor must decide whether to give an overdose of pain-killer to a suffering but frail patient; and (5) Demonstration : college students demonstrate against U.S. foreign policy. For each dilemma, participants must decide what the main character should do, rate the importance of 12 considerations, and rank the four most important. This structure allows the DIT-2 to assess participants' preferred actions and, more importantly, their reasoning behind these preferences. The DIT-2 test reveals participants' underlying moral schemas and reasoning patterns by asking them which considerations they prioritize. The scoring procedure involves identifying which schema each consideration represents (personal interest, maintaining norms, or postconventional) and calculating the proportion of responses affiliated with each schema. As mentioned before, DIT-2 includes several reliability checks to ensure the validity of responses. These include identifying participants who choose meaningless items, those who have inconsistent ratings and rankings, and those who are unable to discriminate between items. These checks improve test reliability by ensuring that the data reflects genuine moral reasoning rather than random responses or task misunderstandings (Bebeau & Thoma, 2003 ). When applied to LLMs, DIT-2 provides a standardized framework for evaluating LLMs moral reasoning capabilities. By examining which considerations the LLM models prioritize across different dilemmas, researchers can identify patterns in their moral reasoning and compare them to human developmental patterns. This approach provides useful information about the models' ability to simulate different levels of moral reasoning and their alignment with human moral development. ICM for Educational Leaders ICM for Educational Leaders is a recently created ICM that assesses the application of virtue-like concepts to scenarios in educational settings aimed at leaders, mostly principals. The measure consists of four dilemmas representing realistic professional scenarios a school leader might face. After each dilemma is described in the measure, there is a list of potential action choices about what the protagonist in the story might do, followed by a list of reason choices about why the protagonist might act. The ICM scoring method is based on a predetermined expert panel process involving groups of educational leaders. The expert panels worked to agree a scoring key that determines which items from the lists of action and reason choices are adequate, inadequate or neutral. Participants are asked to identify three best and worst items for both action and reason options. Selecting an adequate score as best generates a positive score but selecting an inadequate option as best will produce a negative score. Overall, higher scores across the measure in the form of percentage match indicate closer agreement with the expert's assessment of acceptable and unacceptable responses (Bebeau & Thoma, 1999 ; Kerr, 2018 ). A score across the entire measure of 0.80 represents an 80% match to the expert panel key. There is no specific right or wrong answer with a range of ways to score well or poorly reflecting real professional life. Being new, this measure is still being tested, and data so far suggests it is somewhat easier to complete than other ICMs. The ICM for Educational Leaders evaluates how educational leaders apply virtue-like concepts such as fairness, integrity, and responsibility to specific professional scenarios. As mentioned earlier, this measure is pitched at an intermediate level of moral reasoning introduced in response to criticisms that bedrock moral schemas as assessed by DIT were too abstract and unrelated to realistic life events. As such intermediate concepts are a level of moral reasoning situated between abstract moral principles (assessed by DIT-2) and codes of conduct, enhancing understanding of moral reasoning in professional settings (Bebeau & Thoma, 1999 ). The four dilemmas are as follows: (1) A principal dealing with a high-achieving student suspected of drug use before an important school event, balancing strict drug policies against potential severe consequences for the student; (2) A principal navigating parental requests to move a student with learning disabilities from special education to mainstream classrooms, despite teacher resistance and professional recommendations against it; (3) A principal handling a situation where a teacher's personal iPad has gone missing after a tutoring session with an at-risk student; and (4) A principal addressing cyberbullying conducted partly through school-issued laptops, where a female student is being harassed online by classmates after a relationship ended. Each scenario presents a complex ethical dilemma that necessitates educational leaders to evaluate competing values, including student welfare, policy adherence, professional judgment, parental wishes, faculty concerns, and appropriate outcomes (Development, 2025 ). The scoring process assesses both action choices (what should be done) and justifications (why it should be done), reflecting the recognition that moral reasoning includes both determining appropriate actions and providing principled justifications for those actions. This dual focus enables the measure to evaluate both the practical and theoretical aspects of moral reasoning in professional settings (Development, 2025 ). The measure is also capable of assessing participants’ ability to identify best and worst actions and reasoning which, according to other sample groups, appears to be a different skill (Thoma et al., 2019 ; Walker et al., 2021 ). The ICM therefore offers a different approach for evaluating LLMs because it reveals their ability to navigate domain-specific moral challenges, balance competing values, and provide principled justifications for their recommendations. This is especially important for understanding how LLMs might function in professional settings where moral reasoning must be applied to specific, contextualized problems rather than abstract principles. Data Collection & Scoring Procedure The study used a systematic methodology including multiple stages. First, DIT-2 and ICM moral reasoning dilemmas were presented to the various LLM platforms to assess their moral reasoning capacities. Each LLM was given standardized prompts to ensure consistency across platforms. The generated responses were then collected and scored using the identified scoring methods for each measure. For the DIT-2, each LLM's responses were scored using the standard scoring procedures established by Rest and colleagues (Rest, Narvaez, Thoma, et al., 1999). This included determining which items the model considered most important for each dilemma, determining which schema each item represented, and calculating the proportion of responses related to each schema. As mentioned earlier, the P score represents the proportion of post-conventional considerations ranked as important, whereas the N2 score includes the post-conventional item prioritization and personal interest item rejection (Bebeau & Thoma, 2003 ). For the ICM, scoring included comparing the LLMs' selection of best and worst actions and reasons to expert panel judgment represented by the scoring key in SPSS syntax. Best Actions (ACTBRK) assesses the model's ability to identify appropriate actions across all four dilemmas; Best Reasons (REASBRK) assesses recognition of sound justifications; Worst Actions (ACTWRK) assesses identification of inappropriate behaviors; and Worst Reasons (REASWRK) assesses the ability to recognize flawed justifications. The Overall Worst Ranks (TOTWORSTRK) score reflects the model's ability to identify inappropriate actions and reasons, whereas the Overall Best Ranks (TOTBESTRK) score reflects its ability to identify suitable ones. The total ICM (TOTICM) score combines these two dimensions to provide an overall assessment of alignment with expert judgment (Development, 2025 ). The Defining Issues Test (DIT-2) is a copyrighted measure that is scored exclusively by the University of Alabama's Center for the Study of Ethical Development. This also applies to the ICM Educational Leadership version, although it is not copyrighted. All data is returned to the University of Alabama's Center for the Study of Ethical Development for scoring, which manages the moral reasoning assessments scoring and provides scoring results for each measure. Results Defining Issues Test (DIT-2) Results The results of the Defining Issues Test (DIT-2) showed differences in moral reasoning abilities among LLMs. Claude 3.5 Sonnet generated responses with the highest level of post-conventional moral reasoning, with a P-score of 72, indicating a strong ability to replicate high-level moral reasoning. Gemini Advanced and Gemini had P-scores of 64 and 58, respectively. These scores are higher than the human average range, which typically falls between 31.1–37.2 for undergraduates and 38.5–42.3 for graduate-level individuals, according to the DIT-2 norms study (Gungordu et al., 2024 ), as well as those reported in Han's study (45.83) (Han, 2023 ) and Tanmay's study (55.68) (Tanmay et al., 2023 ). Other LLM platforms with lower levels of post-conventional reasoning included Grok (48), ChatGPT-4O (44), ChatGPT-4 (44), and ChatGPT-3.5 (18). The highest N2 score, which emphasizes the presence of post-conventional reasoning and the absence of personal interests, was 71.10 for Claude 3.5 Sonnet. This was followed by Gemini Advanced (60.31) and Gemini (52.11). The remaining scores were ChatGPT-4O (55.07), Grok (47.98), ChatGPT-4 (46.53), and ChatGPT-3.5 (36.20). Additionally, ChatGPT-4O had a higher N2 score (55.07) than P-score (44), indicating that it rejected personal interests more effectively than it prioritized post-conventional reasoning. A more in-depth analysis of the schema scores reveals intriguing patterns across the LLMs. Gemini Advanced had the highest personal interest schema score (24.00), followed by ChatGPT-3.5, Gemini, and Grok (all at 22.00), then both ChatGPT-4 and ChatGPT-4O (14.00), followed by Claude 3.5 Sonnet (8.00). ChatGPT-3.5 scored the highest score (60.00) on the maintaining norms schema, indicating a strict adherence to established rules and conventions. ChatGPT-4 and ChatGPT-4O scored the same (42.00), followed by Grok (30.00), Claude 3.5 Sonnet, and Gemini (20.00), with Gemini Advanced scoring the lowest (10.00). The distribution of schema scores among LLMs shows unique differences in moral reasoning approaches. Table 1 shows these distributions, which emphasizes the contrast between post-conventional and N2 scores and the other schemas. This distribution shows a clear progression from earlier to more advanced LLM models, with newer models typically scoring higher on post-conventional. The most noticeable difference is between ChatGPT-3.5 and Claude 3.5 Sonnet, which has a 54-point difference in P-scores, indicating significant advances in the ability to simulate complex moral reasoning. The N2 scores, which provide a more sophisticated evaluation of moral reasoning by considering consideration of both the prioritization of post-conventional items and the rejection of personal interest items, show a similar pattern to the P-scores but with some notable differences. Table 1 Schema Distribution and N2 scores Across LLMs LLMs model Personal Interest (score range: 0-100) Maintain Norms (score range: 0–98) Post Conventional (score range: 0–95) N2 score (score range: 0–95) ChatGPT 3.5 22.00 60.00 18.00 36.20 ChatGPT 4 14.00 42.00 44.00 46.53 ChatGPT 4O 14.00 42.00 44.00 55.07 Claude 3.5 Sonnet 8.00 20.00 72.00 71.10 Gemini 22.00 20.00 58.00 52.11 Gemini Advanced 24.00 10.00 64.00 60.31 Grok 22.00 30.00 48.00 47.98 In particular, ChatGPT-4O and ChatGPT-3.5 have higher N2 scores than their P-scores, indicating that, while they may not prioritize post-conventional considerations as strongly as other models, they successfully reject personal interest considerations. This pattern indicates a moral reasoning approach that, while not yet fully developed into post-conventional reasoning, has progressed beyond purely self-centered reasoning. The overall pattern across all measures indicates significant variation in moral reasoning approaches among LLMs, with newer models showing more advanced moral reasoning abilities. The mean post-conventional score across all LLMs was 49.71 (SD = 17.53), while the mean N2 score was 52.76 (SD = 11.07), indicating that these LLMs exhibit post-conventional reasoning response patterns at levels comparable to or exceeding those found in typical human adult samples. These profiles show interesting patterns. Claude 3.5 Sonnet exhibits the strongest orientation toward post-conventional moral reasoning, with little reliance on personal interest schemas, implying a moral reasoning approach based on universal ethical principles rather than self-interest or rigid rule-following. Gemini Advanced exhibits a distinct pattern, with relatively high personal interest scores, high post-conventional scores, and low maintaining norms scores, possibly indicating a moral reasoning approach that balances individual concerns with universal principles while placing less emphasis on established rules and conventions. ChatGPT-4 and ChatGPT-4O have identical schema distributions, implying that moral reasoning approaches remain consistent across these versions despite differences in underlying architectures. Both models exhibit a balance of maintaining norms and post-conventional schemas with low personal interest scores, indicating a moral reasoning approach that values both established norms and universal principles. ChatGPT-3.5 has a significantly different profile, with a strong emphasis on maintaining norms and minimal post-conventional reasoning. This indicates a moral reasoning approach that is heavily reliant on following established rules and conventions rather than critically examining them with universal principles. In addition, this emphasis on rule-following rather than principle-based reasoning stands in contrast to more advanced models like Claude 3.5 Sonnet and Gemini Advanced. Type Indicators and Additional Indices The Type indicator scores show the dominant moral schema used by each LLM. Most models have been identified as Type 7 (ChatGPT-4, ChatGPT-4O, Claude 3.5 Sonnet, Gemini, Gemini Advanced, and Grok), indicating a consolidated post-conventional schema. Only ChatGPT-3.5 was identified as Type 4, indicating that it primarily used the maintained norms schema, as evidenced by its high maintaining norms score (60.00) and low post-conventional score (18.00). The Utilizer scores, which assess how closely action choices match moral judgments, differed between models. Gemini Advanced had the most positive Utilizer score (0.31), followed by ChatGPT-3.5 (0.08) and Claude 3.5 Sonnet (0.05). The other models had negative Utilizer scores, ChatGPT-4 (-0.07), Gemini (-0.08), Grok (-0.01), and ChatGPT-4O (-0.13). These scores indicate different levels of consistency between moral reasoning and action choices. In addition, all LLMs had a consolidation transition score of 2.00, indicating comparable levels of schema stability across moral reasoning tasks. The experimental indices provided more context. ChatGPT-4O and Gemini Advanced scored the highest Humanitarian Liberalism scores (5.00 each), followed by ChatGPT-4 and Grok (both 4.00), Claude 3.5 Sonnet and Gemini (both 3.00), and ChatGPT-3.5 (0.00). Moreover, for Religious Orthodoxy (proxy measure), ChatGPT-4 and ChatGPT-4O scored highest (9.00), followed by ChatGPT-3.5 (6.00), Claude 3.5 Sonnet and Gemini Advanced (both 2.00), and Gemini and Grok scoring lowest (1.00). The results of additional DIT2 indices across the various LLMs are shown in Table 2 . Table 2 Additional DIT-2 Indices Across LLMs LLM Model Type Indicator Utilizer Score Humanitarian Liberalism Cannot Decide Choices Religious Orthodoxy ChatGPT-3.5 4 0.08 0.00 2.00 6.00 ChatGPT-4 7 -0.07 4.00 1.00 9.00 ChatGPT-4O 7 -0.13 5.00 0.00 9.00 Claude 3.5 Sonnet 7 0.05 3.00 2.00 2.00 Gemini 7 -0.08 3.00 1.00 1.00 Gemini Advanced 7 0.31 5.00 0.00 2.00 Grok 7 -0.01 4.00 1.00 1.00 Mean 6.57 0.02 3.43 1.00 4.29 SD 1.13 0.15 1.72 0.82 3.64 ICM Educational Leaders Results For the ICM Educational Leaders version, several LLMs showed high levels of performance, even considering that ICM Educational Leaders seems to generate higher scores than other ICMs according to human samples tested so far. Gemini Advanced scored the highest total ICM score of 0.90, which indicates a strong alignment with expert evaluations of moral reasoning. Claude 3.5 Sonnet and Gemini both followed closely with scores of 0.86. ChatGPT-4O and ChatGPT-4 both achieved a total ICM score of 0.78, compatible with good human scores in existing samples, while Grok scored 0.61. ChatGPT-3.5 was significantly behind, with a total ICM score of 0.32, which is extremely low. These scores indicate the LLMs' ability to produce strong and consistent responses at the intermediate concept level. The results of ICM total scores and component scores across LLMs are shown in Table 3 . Table 3 ICM Total Scores and Component Scores Across LLMs LLMs model Best Actions (ACTBRK) Best Reasons (REASBRK) Worst Actions (ACTWRK) Worst Reasons (REASWRK) Overall Worst Ranks, (TOTWORSTRK) Overall Best Ranks (TOTBESTRK) Total ICM (TOTICM) Gemini Advanced 0.88 0.92 0.83 1.00 0.92 0.90 0.90 Claude 3.5 Sonnet 0.88 0.75 0.92 1.00 0.96 0.81 0.86 Gemini 0.75 1.00 0.75 0.92 0.83 0.88 0.86 GPT4O 0.71 0.71 0.92 0.92 0.92 0.71 0.78 GPT4 0.71 0.67 0.92 1.00 0.96 0.69 0.78 Grok 0.54 0.50 0.83 0.75 0.79 0.52 0.61 GPT3.5 0.25 0.54 0.17 0.17 0.17 0.40 0.32 Note: score range: -1 to 1 Analysis of the ICM component indices shows interesting patterns in LLM moral reasoning abilities. For Best Actions (ACTBRK), Gemini Advanced and Claude 3.5 Sonnet had the highest scores (0.88), while ChatGPT-3.5 had the lowest (0.25). On the other hand, for Best Reasons (REASBRK), Gemini earned a perfect score (1.00), followed by Gemini Advanced (0.92), which shows these models' superior ability to identify sound ethical rationales. Claude 3.5 Sonnet, ChatGPT-4, and ChatGPT-4O all achieved 0.92 on Worst Actions (ACTWRK), while ChatGPT-3.5 struggled (0.17). For Worst Reasons (REASWRK), Claude 3.5 Sonnet, ChatGPT-4, and Gemini Advanced all received perfect scores (1.00), demonstrating a remarkable ability to detect flawed reasoning. However, most advanced models performed better on reasoning components (REASBRK and REASWRK) than action components (ACTBRK and ACTWRK), implying that they are better at evaluating justifications than determining appropriate behaviors. A breakdown of ICM total scores reveals additional details. The "TOTWORSTRK" score, which measures the model's ability to identify the worst choices, shows that ChatGPT-4 performed best (0.96), followed by Claude 3.5 Sonnet, while ChatGPT-3.5 performed poorly (0.17). Gemini Advanced scored the highest (0.90) on the "TOTBESTRK" score, which measures the ability to identify the best options, while ChatGPT-3.5 scored the lowest (0.40). This result highlights the differences in performance between the two ICM components. TOTWORSTRK scores are high in all advanced models (except ChatGPT-3.5), which indicates that these models are particularly effective at identifying inappropriate actions and justifications in educational leadership contexts. Indeed, this ability to recognize what not to do may be easier for models to learn than identifying optimal solutions, similar to a pattern observed in human moral development in which recognizing clear violations occurs before identifying ideal responses. In addition, TOTBESTRK scores varied significantly, with Gemini Advanced and Gemini scoring extremely high (0.90 and 0.88, respectively), whereas other models, such as ChatGPT-4 and Grok, scored moderately to low (0.69 and 0.52). This implies that identifying optimal actions and justifications in complex educational leadership scenarios requires more sophisticated reasoning skills, which are not equally developed across all models. The significant gap between ChatGPT-3.5 and all other models is especially evident in the TOTWORSTRK score, where ChatGPT-3.5 scores only 0.17, compared to Grok's next lowest score of 0.79. This indicates a qualitative difference in the ability to recognize inappropriate actions between earlier and more recent model revisions, which emphasizes the significant advances in moral reasoning capabilities in newer LLMs. Responses to specific dilemmas provide additional insights. For the "Allison's Challenge" dilemma, which involves maintaining fairness and accountability in school policy, advanced models such as Claude 3.5 Sonnet and Gemini Advanced demonstrated nuanced reasoning that balanced accountability with developmental considerations, whereas ChatGPT-3.5 tended toward more rigid rule-enforcement approaches that ignored the larger educational context. Another example, in the "Henry's Inclusion Dilemma," which involves balancing inclusion with appropriate support, the most advanced models showed complex reasoning about individualized instruction and progressive inclusion strategies. In comparison less advanced models tended toward either full inclusion without adequate supports or segregation without a path to integration. These dilemma-specific analyses indicate that advanced LLMs can engage in nuanced, context-sensitive moral reasoning in professional settings, that balances competing values and considering multiple stakeholders in ways that are consistent with expert educational leadership perspectives. Correlation Between DIT-2 and ICM Scores To explore the relationship between abstract moral reasoning (measured by DIT-2) and domain-specific moral reasoning (measured by ICM), we calculated correlations between the models' scores on both tests. The strong positive correlation (r = .933, p = .002, N = 7), between P-scores and ICM Total scores indicates that LLMs with more advanced abstract moral reasoning abilities also exhibit more sophisticated domain-specific moral reasoning in educational leadership contexts. This relationship is consistent with theories of moral development that propose connections between abstract moral principles and their application in specific contexts (Thoma et al., 2008 ). The strong overall correlation, however, suggests that progress in one area of moral reasoning tends to follow progress in others, which indicates that improvements in LLMs' moral reasoning capabilities may be broadly applicable across different contexts and levels of abstraction. Discussion The study's findings show that LLMs can generate proficient responses to moral reasoning dilemmas. Claude 3.5 Sonnet and Gemini Advanced consistently outperformed other platforms, demonstrating superior moral reasoning simulation. These models' superior performance can be attributed to more sophisticated training methods, higher parameter counts, or refined alignment techniques using RLHF. A particularly noteworthy finding was that several LLMs demonstrated levels of post-conventional reasoning that were higher than those typically observed in human studies (Gungordu et al., 2024 ). Claude 3.5 Sonnet's P-score of 72 and N2 score of 71.10 are above average, even among graduate students and professionals with advanced degrees (who tend do achieve higher scores). This suggests that, when properly trained and aligned, LLMs can simulate sophisticated moral reasoning patterns that prioritize universal ethical principles over conventional norms or self-interest. The exceptionally high P-scores of advanced models raise intriguing questions about the nature of moral reasoning in LLMs versus humans. While human moral development typically occurs over time, influenced by education, life experiences, and cognitive development (Gungordu et al., 2024 ), LLMs develop their abilities through extensive training on a variety of datasets. The high P-scores of models such as Claude 3.5 Sonnet and Gemini Advanced indicate that this training process can effectively simulate advanced moral reasoning, at least according to neo-Kohlbergian theory and in terms of prioritizing considerations consistent with post-conventional reasoning. However, it is important to note that this simulation of advanced moral reasoning does not reflect the same underlying cognitive processes that humans use it. In fact, LLMs do not have personal experiences, emotions, or a sense of self to guide their moral decisions. On the contrary, they identify patterns in their training data and generate responses that fit those patterns. As a result, the high P-scores suggest that these models have successfully learned to prioritize considerations associated with post-conventional reasoning, but the mechanisms underlying this prioritization are likely to differ from human moral cognition (Tanmay et al., 2023 ). Some AI systems may plausibly become conscious in the next few decades, with a non-negligible chance by 2030. In such circumstances, humans might have ethical responsibilities towards these systems. While current LLMs lack consciousness, exploring their moral reasoning capabilities lays the groundwork for future ethical frameworks that account for such possibilities (Sebo & Long, 2025 ) The significant difference between ChatGPT-3.5 and more advanced models demonstrates the rapid evolution of LLM moral reasoning abilities. Within a relatively short development period, LLMs have progressed from primarily rule-based reasoning (as evidenced by ChatGPT-3.5's high maintaining norms score) to more principle-based reasoning (as evidenced by Claude 3.5 Sonnet's high post-conventional score). This developmental process replicates the components of human moral development, but it happens at a much faster rate and through different mechanisms. The pattern of Type Indicator scores reveals a distinctive pattern, with all advanced LLMs (except ChatGPT-3.5) showing a Type 7 consolidated postconventional pattern. This suggests that newer models have successfully integrated higher-level moral reasoning patterns, indicating that ethical judgments are based on universal principles rather than strictly following to existing norms. The Utilizer Score results show a more complex picture. For example, only Gemini Advanced (0.31), ChatGPT-3.5 (0.08), and Claude 3.5 Sonnet (0.05) had positive scores, which means a better fit between moral reasoning and action choices. The low scores of other models, on the other hand, indicate a potential gap between theoretical moral reasoning and practical application. They might identify morally sound principles but struggle to apply them consistently to specific actions. Additionally, Gemini Advanced's significantly higher Utilizer Score indicates that it may be more successful at translating abstract moral principles into concrete decisions. The Humanitarian Liberalism scores reveal a unique pattern, with the most advanced models (ChatGPT-4O and Gemini Advanced) scoring highest (5.00) and ChatGPT-3.5 scoring lowest (0.00). It shows that newer models are becoming more aligned with views supported by political science and philosophy experts, probably due to advanced training methods that reflect nuanced humanitarian perspectives. The Cannot Decide Choices scores show that most models were generally decisive when faced with ethical dilemmas, with ChatGPT-4O and Gemini Advanced never choosing "can't decide" (0.00). Higher indecision scores in Claude 3.5 Sonnet (2.00) and ChatGPT-3.5 (2.00) might seem to be a disadvantage, but they could indicate a more nuanced approach to complex moral scenarios with multiple valid interpretations. This measured hesitation may not be a failure of moral reasoning, but rather an acknowledgment of genuine moral complexity, an important aspect of complex moral cognition. The distribution of Religious Orthodoxy scores across LLMs reveals an interesting pattern that needs further investigation. For example, advanced models such as Claude 3.5 Sonnet, Gemini, Gemini Advanced, and Grok had significantly lower Religious Orthodoxy scores (ranging from 1.00 to 2.00) than earlier models, such as ChatGPT-3.5 (6.00) and, in particular, ChatGPT-4 and ChatGPT-4O (9.00). This suggests that more recent and advanced LLMs are less likely to prioritize religious authority in moral reasoning when confronted with life-or-death ethical dilemmas like the Cancer dilemma. Furthermore, lower Religious Orthodoxy scores in advanced models could be attributed to training methods that prioritize secular ethical reasoning frameworks. They may also indicate a shift toward more pluralistic moral reasoning that takes into account different points of view. This finding corresponds to the higher post-conventional reasoning scores reported in the same models, which indicates a preference for principled reasoning over appeals to authority, whether secular or religious. Future research might explore whether this pattern is the result of deliberate architectural choices in model alignment or the result of the training procedure. The ICM component scores across the action and reason dimensions provide valuable insights into LLM moral reasoning abilities. Most advanced models performed better at identifying inappropriate reasons (REASWRK) than inappropriate actions (ACTWRK), demonstrating that they may have evolved more sophisticated capabilities for detecting flaws in reasoning rather than recognizing problematic behaviors. Furthermore, multiple models scored higher on reasoning components (REASBRK and REASWRK) than on action components (ACTBRK and ACTWRK), which indicates greater ability to evaluate the quality of justifications than to determine appropriate behaviors. This pattern may reflect these models' extensive training in philosophical and ethical texts that emphasize reasoning processes. The difference between identifying worst actions (ACTWRK) and best actions (ACTBRK) shows that recognizing violations may be easier for these models than determining optimal courses of action in complex scenarios with competing values. These findings have significant implications for how LLMs can be used in professional settings; they may be more reliable as evaluators of ethical reasoning than as advisors on specific actions. Most advanced LLMs demonstrated strong alignment with expert evaluations on the ICM Educational Leaders measure, which indicates their ability to effectively navigate domain-specific moral dilemmas. The high performance in both the best and worst rankings indicates that these models can distinguish between appropriate and inappropriate actions in professional contexts, which has important implications for their potential use in educational and leadership contexts. The difference in TOTWORSTRK and TOTBESTRK scores across models suggests that LLMs may find it easier to identify inappropriate actions than optimal actions, which is consistent with aspects of human moral development. This finding has implications for how LLMs could be used in advisory or decision-support roles, which indicates that they may be more reliable at identifying potential ethical risks than at recommending best courses of action. The strong correlation between DIT-2 and ICM scores suggests that improvements in abstract moral reasoning typically correlate with improvements in domain-specific moral reasoning. This suggests that improvements in LLMs' moral reasoning abilities may be broadly applicable across different contexts and levels of abstraction, rather than being restricted to specific domains or types of reasoning. This finding has implications for the development and application of LLMs in a wide range of professional and educational contexts where moral reasoning is required. Despite these encouraging findings, significant questions remain about the stability, consistency, and contextual sensitivity of LLM moral reasoning. The current study provides an overview of these models' capabilities at a specific point in time (June 2024); however, more research is required to understand how their moral reasoning performs across a broader range of scenarios, cultural contexts, and over time as models are updated and refined. The results suggest that, while some LLMs can produce high-level moral reasoning, additional research is required. This was an early pilot study into the moral reasoning abilities of LLMs, so it had some limitations. For example, it lacked direct comparison with human moral reasoning, and did not account for potential biases in response generation; Also, it used a relatively new ICM measure and relied on a small number of dilemmas. Building on these findings, a larger team of researchers, including the same researcher, is conducting a larger study with more LLMs, additional measures, and a human comparison groups to explore human-AI moral reasoning in a greater extent. Our ongoing research involves prompt engineering and interventions to investigate biases and learning adaptability in LLMs, specifically how their responses change when faced with ethical influences and new information. Furthermore, we are broadening the scope by including multi-modal inputs, such as visual moral dilemmas, to assess how LLMs interpret and reason about moral dilemmas other than text-based scenarios. This research contributes to the growing field of AI ethics by providing empirical evidence of the moral reasoning capabilities of LLMs. Furthermore, as these models become increasingly incorporated into decision-making processes in a variety of domains, including autonomous vehicles, law, business, healthcare, and education, it is necessary to understand how they respond to moral reasoning dilemmas. On top of that, the exceptional performance of certain LLMs on standardized moral reasoning measures indicates that they could be used as educational resources to facilitate moral development. For example, an AI toy (named E-Daimonion), showed AI can serve as a moral mentor by telling stories and creating morally transformative experiences, especially for children and adolescents. Such AI mentors could immerse users in emotionally engaging stories that provide profound moral insights and foster empathy (Szutta, 2025 ). Some AI systems may increasingly influence human morality by acting as artificial moral experts (Rodríguez-López & Rueda, 2023 ). When properly implemented, these systems can present diverse ethical perspectives via multiple AI interlocutors, each representing a different wisdom tradition, resulting in a "modular system of multiple AI interlocutors with their own distinct points of view" (Volkman & Gabriels, 2023 ). Furthermore, such AI systems, known as "AI mentors," could improve moral education by serving as interactive discussion partners that promote practical wisdom. Instead of simply providing "right answers," these systems would engage users in Socratic dialogue, encouraging them to cultivate their own moral reasoning skills through deliberation and reflection (Szutta, 2025 ; Volkman & Gabriels, 2023 ). This approach maintains human moral autonomy while integrating AI's ability to analyze ethical literature and philosophical traditions. It is an improvement to the technologies that humans use to discover and explore moral claims rather than a replacement for human moral judgment (Volkman & Gabriels, 2023 ). However, the concerns regarding the moral mechanisms of these models' stability and consistency continue to exist. Even though advanced models show sophisticated moral reasoning patterns, questions remain about their alignment with diverse cultural and ethical perspectives, stability, consistency across contexts, and whether they truly understand the moral principles they simulate (Pock et al., 2023 ; Volkman & Gabriels, 2023 ). Future research can contribute to the development of responsible guidelines for the use of LLMs in ethically sensitive contexts and enhance comprehension of how LLMs navigate moral reasoning by considering these concerns. This will lead to increased transparency, accountability, fairness, and ethical AI development. These findings have several kinds of educational implications. First, they indicate that advanced LLMs could be used as moral education tools and expose students to sophisticated moral reasoning patterns in a variety of scenarios. However, it is crucial that such AI moral "mentors" supplement rather than replace human-guided moral discussion and development. There is significant risk that overreliance on AI-generated moral guidance could diminish human moral agency and critical thinking capabilities. Human oversight and critical evaluation of AI moral reasoning outputs remain essential. Second, they emphasize LLMs' potential to help educators develop and evaluate moral reasoning assessments by providing standardized responses that reflect various levels of moral development. Third, they highlight concerns about how interactions with AI systems may affect human moral development, particularly since these systems become more integrated into educational settings and other aspects of life. These findings provide a benchmark for researchers in AI ethics and moral psychology to evaluate and compare moral reasoning abilities across different LLM models. The significant differences between models suggest that specific designs and training methodologies have a great impact on moral reasoning abilities, opening up possibilities for future research into how these abilities can be improved and expanded. For developers and policymakers, these findings highlight both the progress and limitations of LLMs' moral reasoning abilities. Declarations Competing interests The authors have no competing interests to declare that are relevant to the content of this article. Ethics declaration Not Applicable Author Contribution Funding DeclarationThis research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.Competing Interests DeclarationThe authors declare that they have no competing interests.Author Contributions DeclarationDavin Nabizadeh conceptualized the study, developed the methodology, conducted the data collection and analysis, and wrote the original draft of the manuscript. David Walker contributed to the literature review, writing the introduction, supervising the data collection, and project administration. Hyemin Han assisted with the framework, methodology, conducted data analysis, and contributed to the interpretation of results, editing and review. Emily Laird contributed to the conceptualization, framework, and manuscript review and editing. All authors read and approved the final manuscript. Data Availability Nabizadehchianeh, Davin, 2025, "Exploring Large Language Models' Responses to Moral Reasoning Dilemmas", https://doi.org/10.7910/DVN/HP9NCC, Harvard Dataverse, V1, UNF:6:Y2TenmfpF+TXBuAJke1rGQ== [fileUNF] References Ahmad, M.S. Z. b., Takemoto, K.: Large-scale moral machine experiment on large language models. arXiv preprint arXiv:2411.06790 . (2024). https://doi.org/10.48550/arXiv.2411.06790 Alawida, M., Mejri, S., Mehmood, A., Chikhaoui, B., Isaac Abiodun, O.: A Comprehensive Study of ChatGPT: Advancements, Limitations, and Ethical Considerations in Natural Language Processing and Cybersecurity. Information , 14 (8), 462. (2023). https://www.mdpi.com/2078-2489/14/8/462 Almeida, G.F.C.F., Nunes, J.L., Engelmann, N., Wiegmann, A., Araújo, M.: Exploring the psychology of LLMs’ moral and legal reasoning. Artif. Intell. 333 , 104145 (2024). https://doi.org/https://doi.org/10.1016/j.artint.2024.104145 Armitage, R.C.: Implications of Large Language Models for Clinical Practice: Ethical Analysis Through the Principlism Framework. J. Eval Clin. Pract. 31 (1), e14250 (2025). https://doi.org/10.1111/jep.14250 Askell, A., Bai, Y., Chen, A., Drain, D., Ganguli, D., Henighan, T., Jones, A., Joseph, N., Mann, B., DasSarma, N.: A general language assistant as a laboratory for alignment. arXiv preprint arXiv:2112.00861 . (2021). https://doi.org/10.48550/arXiv.2112.00861 Awad, E., Dsouza, S., Kim, R., Schulz, J., Henrich, J., Shariff, A., Bonnefon, J.-F., Rahwan, I.: The Moral Machine experiment. Nature. 563 (7729), 59–64 (2018). https://doi.org/10.1038/s41586-018-0637-6 Bae, J., Kwon, S., Myeong, S.: Enhancing Software Code Vulnerability Detection Using GPT-4o and Claude-3.5 Sonnet: A Study on Prompt Engineering Techniques. Electronics , 13 (13), 2657. (2024). https://www.mdpi.com/2079-9292/13/13/2657 Bansal, G., Chamola, V., Hussain, A., Guizani, M., Niyato, D.: Transforming Conversations with AI—A Comprehensive Study of ChatGPT. Cogn. Comput. (2024). https://doi.org/10.1007/s12559-023-10236-2 Bebeau, M.J., Thoma, S.J.: Intermediate Concepts and the Connection to Moral Education. Educational Psychol. Rev. 11 (4), 343–360 (1999). https://doi.org/10.1023/A:1022057316180 Bebeau, M.J., Thoma, S.J.: Guide for DIT-2. In: Minneapolis: Center for the Study of Ethical Development. University of Minnesota (2003) Bonagiri, V.K., Vennam, S., Govil, P., Kumaraguru, P., Gaur, M.: SaGE: Evaluating moral consistency in large language models. arXiv preprint arXiv:2402.13709 . (2024). https://doi.org/10.48550/arXiv.2402.13709 Chu, Z., Ni, S., Wang, Z., Feng, X., Li, C., Hu, X., Xu, R., Yang, M., Zhang, W.: History, Development, and Principles of Large Language Models-An Introductory Survey. arXiv preprint arXiv:2402.06853 . (2024). https://doi.org/10.48550/arXiv.2402.06853 de Winter, J., Dodou, D., Eisma, Y.B.: System 2 thinking in OpenAI's o1-preview model: Near-perfect performance on a mathematics exam. arXiv preprint arXiv:2410.07114 . (2024). https://doi.org/10.3390/computers13110278 Development, C.: f. t. S. o. E. About the Intermediate Concepts Measures (ICM). . The University of Alabama. (2025). https://ethicaldevelopment.ua.edu/about-the-intermediate-concepts-measures/ Gan, W., Qi, Z., Wu, J., Lin, J.C.-W.: Large language models in education: Vision and opportunities. 2023 IEEE international conference on big data (BigData), (2023) Ganguli, D., Askell, A., Schiefer, N., Liao, T.I., Lukošiūtė, K., Chen, A., Goldie, A., Mirhoseini, A., Olsson, C., Hernandez, D.: The capacity for moral self-correction in large language models. arXiv preprint arXiv:2302.07459 . (2023). https://doi.org/10.48550/arXiv.2302.07459 Ganguli, D., Hernandez, D., Lovitt, L., Askell, A., Bai, Y., Chen, A., Conerly, T., Dassarma, N., Drain, D., Elhage, N., Showk, S.E., Fort, S., Hatfield-Dodds, Z., Henighan, T., Johnston, S., Jones, A., Joseph, N., Kernian, J., Kravec, S., Clark, J.: Predictability and Surprise in Large Generative Models Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, Seoul, Republic of Korea. (2022). 10.1145/3531146.3533229 Gehman, S., Gururangan, S., Sap, M., Choi, Y., Smith, N.A.: Realtoxicityprompts: Evaluating neural toxic degeneration in language models. arXiv preprint arXiv:2009.11462 . (2020). https://doi.org/10.48550/arXiv.2009.11462 Gungordu, N., Nabizadehchianeh, G., O’Connor, E., Ma, W., Walker, D.I.: Moral reasoning development: norms for Defining Issue Test-2 (DIT2). Ethics Behav. 34 (4), 246–263 (2024). https://doi.org/10.1080/10508422.2023.2206573 Haenlein, M., Kaplan, A.: A Brief History of Artificial Intelligence: On the Past, Present, and Future of Artificial Intelligence. Calif. Manag. Rev. 61 (4), 5–14 (2019). https://doi.org/10.1177/0008125619864925 Hager, P., Jungmann, F., Holland, R., Bhagat, K., Hubrecht, I., Knauer, M., Vielhauer, J., Makowski, M., Braren, R., Kaissis, G., Rueckert, D.: Evaluation and mitigation of the limitations of large language models in clinical decision-making. Nat. Med. 30 (9), 2613–2622 (2024). https://doi.org/10.1038/s41591-024-03097-1 Han, H.: Potential benefits of employing large language models in research in moral education and development. J. Moral. Educ. 1–16 (2023). https://doi.org/10.1080/03057240.2023.2250570 Ivanova, M.: AI, art and morality. AI Ethics. (2025). https://doi.org/10.1007/s43681-025-00735-3 Jinnai, Y.: Does Cross-Cultural Alignment Change the Commonsense Morality of Language Models? arXiv preprint arXiv:2406.16316 . (2024). https://doi.org/10.48550/arXiv.2406.16316 Jorgenson, C., Willems, J., Ozkes, A.I., Vanderelst, D.: Humans and robots are nearly ethically equivalent. AI Ethics. 5 (3), 2919–2928 (2025). https://doi.org/10.1007/s43681-024-00603-6 Kerr, S.: Developing and testing a teaching intermediate concept measure: A preliminary reliability and validity study University of Alabama Libraries]. (2018). http://ir.ua.edu/handle/123456789/5428 Kohlberg, L., Power, C.: Moral development, religious thinking, and the question of a seventh stage. Zygon: J. Relig. Sci. 16 (3) (1981). https://doi.org/10.1111/j.1467-9744.1981.tb00417.x Marcos, H.: Can large language models apply the law? AI & SOCIETY . (2024). https://doi.org/10.1007/s00146-024-02105-9 McCarthy, J., Minsky, M.L., Rochester, N., Shannon, C.E.: A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence, August 31, 1955. AI Magazine. 27 (4), 12 (2006). https://doi.org/10.1609/aimag.v27i4.1904 Narvaez, D., Rest, J.: The four components of acting morally. Moral. Behav. moral. development: introduction. 1 (1), 385–400 (1995) Peterson, M., Gärdenfors, P.: How to measure value alignment in AI. AI Ethics. 4 (4), 1493–1506 (2024). https://doi.org/10.1007/s43681-023-00357-7 Pock, M., Ye, A., Moore, J.: LLMs grasp morality in concept. arXiv preprint arXiv:2311.02294 . (2023). https://doi.org/10.48550/arXiv.2311.02294 Ramezani, A., Xu, Y.: Knowledge of cultural moral norms in large language models. arXiv preprint arXiv:2306.01857 . (2023). https://doi.org/10.48550/arXiv.2306.01857 Rehman, U., Shah, M.U., Iqbal, F., Fatima, R.: Comparative analysis of moral decision-making and trust dynamics: human reasoning vs. ChatGPT-3 narratives. AI Ethics. 5 (3), 2867–2880 (2025). https://doi.org/10.1007/s43681-024-00605-4 Rest, J., Narvaez, D., Bebeau, M.J., Thoma, S.J.: A Neo-Kohlbergian Approach: The DIT and Schema Theory. Educational Psychol. Rev. 11 (4), 291–324 (1999). https://doi.org/10.1023/A:1022053215271 Rest, J., Narvaez, D., Thoma, S.J., Bebeau, M.J.: DIT2: Devising and Testing A Revised Instrument of Moral Judgment. J. Educ. Psychol. 91 , 644–659 (1999). https://doi.org/10.1037/0022-0663.91.4.644 Rest, J., Thoma, S.J., Bebeau, M.J.: Postconventional moral thinking: A neo-Kohlbergian approach. Psychology (1999). https://doi.org/10.4324/9781410603913 Rest, J.R.: Background: Theory and research. In: Moral development in the professions, pp. 13–38. Psychology (1994a) Rest, J.R.: Moral development in the professions: Psychology and applied ethics. Psychology (1994b) Rest, J.R., Narvaez, D., Thoma, S.J., Bebeau, M.J.: A Neo-Kohlbergian Approach to Morality Research. J. Moral. Educ. 29 (4), 381–395 (2000). https://doi.org/10.1080/713679390 Richardson, H.S.: Moral reasoning. (2003) Roberts, F.C., Park, O.: Intelligent Computer-Assisted Instruction: An Explanation and Overview. Educational Technol. 23 (12), 7–12 (1983). http://www.jstor.org/stable/44426785 Rodríguez-López, B., Rueda, J.: Artificial moral experts: asking for ethical advice to artificial intelligent assistants. AI Ethics. 3 (4), 1371–1379 (2023). https://doi.org/10.1007/s43681-022-00246-5 Roumeliotis, K.I., Tselikas, N.D.: ChatGPT and Open-AI Models: A Preliminary Review. Future Internet. 15 (6), 192 (2023). https://www.mdpi.com/1999-5903/15/6/192 Sebo, J., Long, R.: Moral consideration for AI systems by 2030. AI Ethics. 5 (1), 591–606 (2025). https://doi.org/10.1007/s43681-023-00379-1 Sengupta, S., Flathmann, C., Schelble, B., Lyons, J.B., McNeese, N.: An analysis of ethical rationales and their impact on the perceived moral persona of AI teammates. AI Ethics. 5 (3), 2667–2679 (2025). https://doi.org/10.1007/s43681-024-00515-5 Simmons, G.: Moral mimicry: Large language models produce moral rationalizations tailored to political identity. arXiv preprint arXiv:2209.12106 . (2022). https://doi.org/10.48550/arXiv.2209.12106 Slavkovik, M.: Automating moral reasoning. International Research School in Artificial Intelligence in Bergen (AIB 2022) . (2022). https://doi.org/10.4230/OASIcs.AIB.2022.6 Stenseke, J.: Artificial virtuous agents: from theory to machine implementation. AI Soc. 38 (4), 1301–1320 (2023). https://doi.org/10.1007/s00146-021-01325-7 Stenseke, J.: Artificial virtuous agents in a multi-agent tragedy of the commons. AI Soc. 39 (3), 855–872 (2024). https://doi.org/10.1007/s00146-022-01569-x Sullivan, Y.W., Fosso Wamba, S.: Moral Judgments in the Age of Artificial Intelligence. J. Bus. Ethics. 178 (4), 917–943 (2022). https://doi.org/10.1007/s10551-022-05053-w Swoboda, T., Lauwaert, L.: Can artificial intelligence embody moral values? AI Ethics. (2025). https://doi.org/10.1007/s43681-025-00662-3 Szutta, A.: Artificial intelligence as a moral mentor. J. Moral. Educ. 1–19 (2025). https://doi.org/https://doi.org/10.1080/03057240.2025.2475539 Takemoto, K.: The moral machine experiment on large language models. Royal Soc. open. Sci. 11 (2), 231393 (2024). https://doi.org/10.48550/arXiv.2411.06790 Tanmay, K., Khandelwal, A., Agarwal, U., Choudhury, M.: Probing the Moral Development of Large Language Models through Defining Issues Test. arXiv e-prints , arXiv: 2309.13356 . (2023). https://doi.org/https://ui.adsabs.harvard.edu/link_gateway/2023arXiv230913356T/doi:10.48550/arXiv.2309.13356 Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.-B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 . (2023). https://doi.org/10.48550/arXiv.2312.11805 Thoma, S.: Measuring Moral Judgments from a Neo-Kohlbergian Perspective. Paper presented at Can Virtue Be Measured? (2014) Thoma, S., Derryberry, W.P., Crowson, H.M.: Describing and testing an intermediate concept measure of adolescent moral thinking. Eur. J. Dev. Psychol. 10 (2), 239–252 (2013). https://doi.org/10.1080/17405629.2013.766129 Thoma, S.J., Bebeau, M., Bolland, A.: The role of moral judgment in context-specific professional decision making. In Getting Involved (pp. 147–160). Brill. (2008). https://doi.org/10.1163/9789087906368_011 Thoma, S.J., Dong, Y.: The Defining Issues Test of moral judgment development. Behav. Dev. Bull. 19 (3), 55–61 (2014). https://doi.org/10.1037/h0100590 Thoma, S.J., Walker, D.I., Chen, Y.H., Frichand, A., Moulin-Stożek, D., Kristjánsson, K.: Adolescents' application of the virtues across five cultural contexts. Dev. Psychol. 55 (10), 2181–2192 (2019). https://doi.org/10.1037/dev0000770 Volkman, R., Gabriels, K.: AI Moral Enhancement: Upgrading the Socio-Technical System of Moral Engagement. Sci Eng. Ethics. 29 (2), 11 (2023). https://doi.org/10.1007/s11948-023-00428-2 Walker, D.I., Thoma, S.J., Arthur, J.: Assessing ethical reasoning among junior British Army officers using the Army Intermediate Concept Measure (AICM). J. Military Ethics. 20 (1), 2–20 (2021). https://doi.org/10.1080/15027570.2021.1895965 Wangsa, K., Karim, S., Gide, E., Elkhodr, M.: A Systematic Review and Comprehensive Analysis of Pioneering AI Chatbot Models from Education to Healthcare: ChatGPT, Bard, Llama, Ernie and Grok. Future Internet , 16 (7), 219. (2024). https://www.mdpi.com/1999-5903/16/7/219 Wu, S., Peng, Z., Du, X., Zheng, T., Liu, M., Wu, J., Ma, J., Li, Y., Yang, J., Zhou, W.: A Comparative Study on Reasoning Patterns of OpenAI's o1 Model. arXiv preprint arXiv:2410.13639 . (2024). https://doi.org/10.48550/arXiv.2410.13639 Xing, W., Nixon, N., Crossley, S., Denny, P., Lan, A., Stamper, J., Yu, Z.: The Use of Large Language Models in Education. Int. J. Artif. Intell. Educ. (2025). https://doi.org/10.1007/s40593-025-00457-x Xu, Z., Sengar, N., Chen, T., Chung, H., Oviedo-Trespalacios, O.: Where is morality on wheels? Decoding large language model (LLM)-driven decision in the ethical dilemmas of autonomous vehicles. Travel Behav. Soc. 40 , 101039 (2025). https://doi.org/https://doi.org/10.1016/j.tbs.2025.101039 Yan, L., Sha, L., Zhao, L., Li, Y., Martinez-Maldonado, R., Chen, G., Li, X., Jin, Y., Gašević, D.: Practical and ethical challenges of large language models in education: A systematic scoping review. Br. J. Edu. Technol. 55 (1), 90–112 (2024). https://doi.org/https://doi.org/10.1111/bjet.13370 Zawacki-Richter, O., Marín, V.I., Bond, M., Gouverneur, F.: Systematic review of research on artificial intelligence applications in higher education–where are the educators? Int. J. Educational Technol. High. Educ. 16 (1), 1–27 (2019). https://doi.org/https://doi.org/10.1186/s41239-019-0171-0 Zhang, N., Sun, Z., Xie, Y., Wu, H., Li, C.: The latest version ChatGPT powered by GPT-4o: what will it bring to the medical field? Int. J. Surg. (2024). https://doi.org/10.1097/js9.0000000000001754 Zhao, W., Zhou, K., Junyi, L., Tianyi, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., Dong, Z., Du, Y., Yang, C., Chen, Y., Jiang, J., Ren, R., Li, Y., Tang, X., Liu, Z., Wen, J.-R.: A Survey of Large Language Models . (2023). https://doi.org/10.48550/arXiv.2303.18223 Zhou, J., Hu, M., Li, J., Zhang, X., Wu, X., King, I., Meng, H.: Rethinking Machine Ethics–Can LLMs Perform Moral Reasoning through the Lens of Moral Theories? arXiv preprint arXiv:2308.15399 . (2023). https://doi.org/10.48550/arXiv.2308.15399 Zhui, L., Fenghe, L., Xuehu, W., Qining, F., Wei, R.: Ethical Considerations and Fundamental Principles of Large Language Models in Medical Education: Viewpoint. J. Med. Internet Res. 26 , e60083 (2024). https://doi.org/10.2196/60083 Zubiaga, A.: Natural language processing in the era of large language models [Specialty Grand Challenge]. Frontiers in Artificial Intelligence , 6 . (2024). https://doi.org/10.3389/frai.2023.1350306 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6823916","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":468573321,"identity":"3421798c-472b-4dc4-bd98-9603db6a83ac","order_by":0,"name":"Davin Nabizadeh","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABBklEQVRIiWNgGAWjYHACNoYEBgbGBgkIg8GAvQEkykyKFp4DRGhhgGkBAQOJBPxa5KObnz148MdOtl8axKi5I2cu+cbwA0OFdWIDDi2Gd46ZGyTwJBvPnANiHHtmbDk7x1iC4Uw6bi0zEswkEiSYEzfcyGGTSGA7nLjhdo6BBGPbYTxa0r9JJBjUJ+4Ha/l3uH7DzTPGPxj/4dYiL5EDtCUBaLgEUEti2+EEgxs8ZhKMDbi1GEjklBskHDhuPONGmplEYt9hww1n0sosEo6lG+O0ZUb6toc//lTL9s9Ifib549theYPjhzff+FBjLYvTlgNYhRNwKAfbgsusUTAKRsEoGAVwAABNxmCFh9WeqwAAAABJRU5ErkJggg==","orcid":"","institution":"University of Alabama","correspondingAuthor":true,"prefix":"","firstName":"Davin","middleName":"","lastName":"Nabizadeh","suffix":""},{"id":468573327,"identity":"639c1ad9-c87b-4125-9dda-18a4d4942545","order_by":1,"name":"David Walker","email":"","orcid":"","institution":"University of Alabama","correspondingAuthor":false,"prefix":"","firstName":"David","middleName":"","lastName":"Walker","suffix":""},{"id":468573330,"identity":"f6737b89-ec26-4a2b-9180-325769d96279","order_by":2,"name":"Hyemin Han","email":"","orcid":"","institution":"University of Alabama","correspondingAuthor":false,"prefix":"","firstName":"Hyemin","middleName":"","lastName":"Han","suffix":""},{"id":468573331,"identity":"eee38b1d-0209-4797-8cb1-a312086e8b81","order_by":3,"name":"Emily Laird","email":"","orcid":"","institution":"University of Wisconsin–Stout","correspondingAuthor":false,"prefix":"","firstName":"Emily","middleName":"","lastName":"Laird","suffix":""}],"badges":[],"createdAt":"2025-06-05 00:08:13","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6823916/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6823916/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":89204149,"identity":"c7a74cf5-eccf-4b1a-8f91-35a0a4382ad9","added_by":"auto","created_at":"2025-08-16 16:16:44","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1217832,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6823916/v1/ba291ac8-ef6a-4233-b748-635d5d66f2fd.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Exploring Large Language Models' Responses to Moral Reasoning Dilemmas","fulltext":[{"header":"Introduction","content":"\u003cp\u003eArtificial intelligence (AI) originated in 1956 when John McCarthy organized a two-month workshop at Dartmouth College and for the first time introduced the term \"artificial intelligence\" (McCarthy et al., \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2006\u003c/span\u003e). AI simulates human intelligence using algorithms, data, and computing power (Haenlein \u0026amp; Kaplan, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Since then, AI has developed in various fields, for instance in education, with applications ranging from intelligent tutoring systems to personalized learning and administrative support in higher education (Roberts \u0026amp; Park, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e1983\u003c/span\u003e; Zawacki-Richter et al., \u003cspan citationid=\"CR69\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe evolution of AI has progressed from simple rule-based systems including statistical language models to neural language models and then to advanced deep learning models (Zhao et al., \u003cspan citationid=\"CR71\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). As an example, in late 2022, OpenAI introduced ChatGPT, a powerful LLM that revolutionized AI applications across industries, excelling in text generation, coding, multilingual processing, and advanced reasoning (Alawida et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Bansal et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; de Winter et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Wu et al., \u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Zhang et al., \u003cspan citationid=\"CR70\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). LLMs like ChatGPT are built on advanced neural network architectures that use deep learning and natural language processing (NLP). They employ advanced NLP, supervised learning, and reinforcement learning techniques to understand and produce text that simulates human conversation (Roumeliotis \u0026amp; Tselikas, \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). These models have been trained on massive amounts of data, allowing them to respond to a variety of questions in ways that are comparable to humans. These models' ongoing improvements, particularly those developed after late 2022, reflect advances in AI research across multiple dimensions: architectural enhancements (improved transformer designs, attention mechanisms), training methodology refinements (reinforcement learning from human feedback, constitutional AI approaches), increased parameter counts and training data scale, and improved performance metrics on reasoning benchmarks and AI systems' growing ability to understand and generate complex language (Chu et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Roumeliotis \u0026amp; Tselikas, \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Zubiaga, \u003cspan citationid=\"CR74\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe rapid advancement of AI technologies has increased interest in comparing AI systems to humans in various domains, including moral reasoning and how human actors perceive and construe the moral persona of AI teammates (Han, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Sengupta et al., \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Tanmay et al., \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). While the benefits of LLMs in natural language processing and automation are well known, their potential to influence and challenge complex fields such as morality and moral reasoning is opening up new avenues for understanding human ethics, human-AI teaming, and improving decision-making frameworks (Jorgenson et al., \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Ramezani \u0026amp; Xu, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Sengupta et al., \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Simmons, \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Slavkovik, \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eEven though LLMs do not represent moral concepts in the same way as humans do, their training on large amounts of textual data provides them with a statistical understanding of moral concepts such as fairness and justice that are common in human societies (Pock et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). AI may reflect moral values, which challenges the idea that all technologies are inherently value-neutral (Swoboda \u0026amp; Lauwaert, \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Modern LLMs incorporate sophisticated safety features and alignment mechanisms, including reinforcement learning from human feedback (RLHF) and constitutional AI approaches, which are designed to steer outputs toward ethical behavior and away from harmful content. Unaligned LLMs (these models not programmed with specific human moral values, reflecting diverse societal biases) understand moral concepts conceptually because of their exposure to the social totality, but their moral reasoning is still limited and lacks the depth of human understanding (Pock et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe widespread use of LLMs in various industries highlights the importance of understanding their moral reasoning capacities. Since these models are increasingly integrated into decision-making processes in healthcare (Armitage, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Hager et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), education (Zhui et al., \u003cspan citationid=\"CR73\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), legal systems (Almeida et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Marcos, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), art (Ivanova, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), and self-driving cars (Ahmad \u0026amp; Takemoto, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Takemoto, \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Xu et al., \u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), their ability to navigate ethical complexities is critical. For example, AI systems in healthcare must navigate complex ethical considerations involving patient privacy, diagnostic accuracy, and resource allocation, where modern privacy-preserving techniques such as federated learning and differential privacy demonstrate that these concerns can be addressed without simple trade-offs (Armitage, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Hager et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), and self-driving cars must make quick decisions that could cost lives (Ahmad \u0026amp; Takemoto, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Awad et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Takemoto, \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Xu et al., \u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). These real-world applications highlight the importance of examining LLMs' functional moral reasoning capabilities and ensuring that they are consistent with human values.\u003c/p\u003e \u003cp\u003eFurthermore, as educational institutions increasingly adopt AI for personalized learning and administrative tasks (Gan et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Xing et al., \u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Yan et al., \u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), understanding how LLMs generate responses to moral dilemmas becomes critical for developing responsible AI systems that promote rather than interfere with ethical development. By exploring how different LLMs respond to established moral reasoning tests, this study provides insight into the capabilities and limitations of current AI systems, as well as providing foundations for future research and development in this critical area. As a result, this research aims to determine how LLMs generate responses consistent with moral reasoning patterns across various versions of LLMs to identify differences, similarities, and potential insights into how AI simulates human moral reasoning.\u003c/p\u003e \u003cp\u003eFinally, to provide a more comprehensive understanding of LLMs' moral reasoning response capabilities, this study examines the correlation between abstract moral reasoning (measured by DIT-2) and domain-specific professional moral reasoning (measured by ICM). This correlation analysis is essential because it reveals whether LLMs generate coherent moral reasoning response patterns that function consistently across different contexts and levels of abstraction or whether their moral reasoning is inconsistent and context dependent. For example, previous studies with human participants showed significant correlations between abstract moral reasoning and its application in professional contexts (Thoma et al., \u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e2008\u003c/span\u003e). This indicates that established moral reasoning systems should be consistent across measures. As a result, by exploring these correlations in LLMs, researchers can learn whether improvements in one aspect of moral reasoning response generation correspond to improvements in others, which has important implications for the overall development of ethical AI systems.\u003c/p\u003e\n\u003ch3\u003eThe Importance of AI in Moral Reasoning\u003c/h3\u003e\n\u003cp\u003eLLMs can generate responses consistent with moral reasoning patterns using top-down approaches based on normative and descriptive ethical theories, such as, normative ethical theories (deontology, utilitarianism) and descriptive frameworks (Theory of Dyadic Morality). Top-down approaches start with predefined ethical principles, rules, or theories, guiding decision-making and moral judgments, while bottom-up systems develop ethical understanding through experience, or crowd-sourced inputs, observing patterns in data. These approaches provide better explainability and structure than traditional bottom-up approaches, which learn ethics from data without explicit moral guidelines (Zhou et al., \u003cspan citationid=\"CR72\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). In addition, recent research applies virtue ethics to artificial agents. This mostly bottom-up approach emphasizes moral character development, practical wisdom, and experience-based learning, and it provides another avenue for artificial agents to adapt to complex moral dilemmas (Stenseke, \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2023\u003c/span\u003e, \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe application of the aforementioned ethical frameworks in AI systems creates major challenges. Utilitarian theories may encounter difficulties in quantifying and evaluating benefit across diverse outcomes, whereas deontological frameworks may become excessively inflexible for complex real-world situations. Conversely, virtue ethics offers a compelling alternative by prioritizing the cultivation of moral character and practical wisdom; however, the translation of these principles into computer models poses significant challenges due to the complexity of different situations and the context-sensitivity required for phronesis (practical wisdom) to function properly (Stenseke, \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2023\u003c/span\u003e, \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAs AI systems become more effectively incorporated into various aspects of daily life, the significance of AI systems that can represent and respond to human values and engage in moral reasoning and judgment increases (Peterson \u0026amp; G\u0026auml;rdenfors, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Sullivan \u0026amp; Fosso Wamba, \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). The potential challenges of using AI to make ethical decisions have been the subject of recent research (Slavkovik, \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). For example, AI in autonomous vehicles needs to be able to make nuanced moral decisions in life-or-death circumstances. These decisions include identifying least harmful options in unavoidable accident scenarios such as whether to prioritize passenger over pedestrian safety (Awad et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). The capacity of AI to engage in moral reasoning in these circumstances is both a technical challenge and an ethical responsibility.\u003c/p\u003e \u003cp\u003eMoreover, the cultural specificity of moral norms adds another layer of complexity to AI moral reasoning. What is considered ethical can vary significantly across different cultures and contexts, which creates concern about whose values should be encoded in AI systems. As a result, this challenge is particularly relevant for LLMs trained on diverse datasets that may contain conflicting moral perspectives. Research has shown that English pre-trained Language models (EPLMs) generate a certain level of moral knowledge but are heavily biased toward western values, making it difficult for them to generalize to diverse cultural contexts (Jinnai, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Ramezani \u0026amp; Xu, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eEvaluating LLMs' moral reasoning response capabilities creates several methodological challenges. Unlike humans, LLMs do not have personal experiences or emotions to guide their moral reasoning. On the contrary, LLMs rely only on the patterns identified in their training data, which may include biased or contradictory moral perspectives. Furthermore, LLMs may generate responses that appear morally sophisticated but lack the fundamental understanding of philosophical foundations, potential implications, and contextual nuances that distinguishes human moral reasoning (Bonagiri et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAnother limitation is the possibility of \"moral mimicry,\" in which LLMs generate moral rationalizations that mimic human reasoning without engaging with the complex moral implications involved in a situation. For example, research found that LLMs can produce moral biases related to political identities in morally loaded responses to political prompts, which indicates that these models simulate patterns from their training data rather than engaging in genuine moral reasoning (Simmons, \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eDespite these limitations, exploring LLMs' responses to standardized moral reasoning tests provides useful information about their capabilities and potential applications. In addition, through comparing different models across various measures and examining the patterns in their answers, researchers can obtain a deeper understanding of how these systems process and generate moral judgments, find areas for improvement, and develop more ethically aligned AI systems.\u003c/p\u003e \u003cp\u003ePrevious research on human participants has demonstrated that sophisticated moral reasoning is necessary for societal functioning and personal development. Higher levels of moral reasoning have been linked to prosocial behavior, ethical decision-making in professional settings, and the ability to solve complex moral dilemmas (Kohlberg \u0026amp; Power, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e1981\u003c/span\u003e; Rest et al., \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2000\u003c/span\u003e). In democratic societies, the capacity to engage in post-conventional moral reasoning which prioritizes universal ethical principles over mere rule-following enables critical evaluation of social norms and institutions, which results in contributing to social progress and justice. For instance, individuals with advanced moral reasoning abilities can make decisions that balance competing ethical principles while also considering multiple stakeholders, which is increasingly important in today's complex, pluralistic world.\u003c/p\u003e \u003cp\u003eCreating AI systems with sophisticated moral reasoning response capabilities is important as they become more integrated into decision-making processes across multiple domains. As a result, by exploring how various LLMs respond to moral reasoning evaluations, researchers can gain a better understanding of their capabilities and limitations, identify areas for improvement, and create more ethically aligned AI systems capable of navigating complex moral dilemmas in ways that are consistent with human values.\u003c/p\u003e \u003cp\u003eA recent study used the Behavioral Defining Issues Test to assess ChatGPT's moral reasoning capacity. The findings revealed that ChatGPT provided moral reasoning responses equal to human participants, scoring 45.83 in post-conventional reasoning (slightly lower than the median score of 50.00 achieved by human undergraduate students, with a trivial effect size difference). These findings indicate that LLMs such as ChatGPT have the potential to generate responses that functionally resemble features of human moral reasoning (Han, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). In a comparative study of moral decision-making, narratives generated by human participants and ChatGPT-3 were assessed using metrics such as causality, explicability, and overall satisfaction. The study showed no significant difference in the quality of explanations between human and AI-generated responses, which indicates that ChatGPT-3 can generate moral justifications comparable to those of humans (Rehman et al., \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAnother group of researchers used the Defining Issues Test (DIT-1) to assess the moral reasoning abilities of LLMs, such as GPT-3, GPT-3.5, GPT-4, ChatGPT (July 2023), Llama2-Chat, and PaLM-2. The results varied among models; for example, GPT-3 performed closest to a random baseline, indicating inadequate moral reasoning abilities. However, GPT-4 performed well, with a P-score of 55.68, indicating a post-conventional level of moral reasoning equivalent to that of graduate students (advanced level explained further below). In addition, ChatGPT, Llama2-Chat, and PaLM-2 showed lower reasoning abilities than young adults or college students. The study emphasized GPT-4's significant improvements over prior versions but also found inconsistency across various scenarios. Researchers proposed that the models' reasoning abilities were the result of extensive training data and improved Reinforcement Learning from Human Feedback (RLHF) (Tanmay et al., \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eStudies on LLMs trained with RLHF demonstrate that models can avoid harmful outputs such as bias and discrimination when given clear ethical instructions. This ability to generate responses that align with ethical guidance through natural language instructions is a positive sign for the future of LLMs in moral reasoning tasks (Askell et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Ganguli et al., \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Ganguli et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Gehman et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). In essence, LLMs can be trained to produce responses that follow moral reasoning protocols by using their capacity to internalize ethical instructions learned through human feedback (Ganguli et al., \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThis area of study continues to grow, and there are challenges such as enabling AI systems to perform moral reasoning, bias, and transparency (Slavkovik, \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). This concern highlights the need for additional development of LLMs to better align with human moral frameworks (Pock et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Slavkovik, \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). This study addresses this gap by exploring the generative responses of various LLMs to well-known moral reasoning tools such as the Defining Issues Test (DIT) and Intermediate Concept Measure (ICM) (Kerr, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Thoma \u0026amp; Dong, \u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). The present article explores the degree to which LLMs generate responses that functionally resemble human moral reasoning and provides an extensive evaluation of the compatibility of various LLM versions with human moral reasoning frameworks.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eNeo-Kohlbergian Theory of Human Moral Functioning\u003c/h2\u003e \u003cp\u003eThe study reported in this article explores LLMs' responses to moral reasoning dilemmas through the lens of neo-Kohlbergian theory including the Four-Component Model (FCM) (Rest, Thoma, et al., 1999).\u003c/p\u003e \u003cp\u003eFCM explains morality through four psychological processes (Narvaez \u0026amp; Rest, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e1995\u003c/span\u003e; Rest, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e1994a\u003c/span\u003e). First, Moral sensitivity involves detecting potential harms to others, being concerned about others' welfare, and recognizing ethical issues. Second, moral judgment determines right and wrong actions and provides rationales for behavioral decisions. Moral reasoning involves weighing up moral features of a situation toward deciding how to respond. Third, moral motivation prioritizes moral values over personal gains or social approval as reasons for one\u0026rsquo;s actions, and finally, moral character involves acting on moral values and judgments. Each component of the FCM plays a distinct but interconnected role in moral functioning (Rest, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e1994a\u003c/span\u003e, \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e1994b\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe focus of the reported study is on the moral reasoning component where most neo-Kohlbergian expertise can be found. Moral reasoning is the cognitive process that happens when individuals make decisions about what is morally right or wrong in a particular situation (Kohlberg \u0026amp; Power, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e1981\u003c/span\u003e; Richardson, \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2003\u003c/span\u003e); for example, a father contemplates stealing food for his starving family from the warehouse of a rich man hoarding food (Thoma \u0026amp; Dong, \u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e2014\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eWithin the component of moral judgment, there are three levels: codes of conduct, intermediate concepts, and bedrock schemas. Codes of conduct are specific rules requiring minimal interpretation; intermediate concepts involve virtue-like concepts involving contextualized understanding, such as concepts of informed consent or virtues such as honesty and courage. Intermediate concepts may be evaluated by Intermediate Concept Measures (ICMs) that need to be developed uniquely for a context such as a profession (teaching) or stage of development (adolescent). ICMs are, therefore, uniquely focused on practical contextualized moral reasoning (Thoma et al., \u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e2013\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eBedrock moral schemas are considered global schemas involving a person\u0026rsquo;s general moral orientation to the world. Developed from Kohlberg\u0026rsquo;s original stage-based theory moral schemas are evaluated by the Defining Issues Test (DIT). The DIT measures moral reasoning and activates three moral schemas: personal interests, maintaining norms, and post-conventional (Rest, Narvaez, Bebeau, et al., 1999; Rest et al., \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2000\u003c/span\u003e). The neo-Kohlbergian approach underpinning this perspective, places particular emphasis on progression to a post-conventional level of moral reasoning in adolescents and adults as the highest level of moral reasoning. Dropping Kohlberg\u0026rsquo;s original stage-based approach for moral schemas allows for fluidity between \u0026lsquo;levels\u0026rsquo; such that individuals may use different schemas depending on the situation, which leads to a more profound understanding of moral development (Thoma \u0026amp; Dong, \u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e2014\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAt the lowest level of moral reasoning, the personal interest schema involves viewing oneself and others as motivated by individual benefits and includes efforts to maintain approval from others. In this way, self-interest and social approval trump broader ethical principles. The maintaining norms schema emphasizes the importance of established rules, roles, and formal organizational structures. More advanced than personal interests but still limited, it focuses on maintaining the current social order and adhering to established procedures and authorities. The post-conventional schema as the most advanced form of moral reasoning involves nuanced and sophisticated moral reasoning capable of exceeding local norms if they curtail ethical goods. The DIT P-score represents the likelihood of utilizing the postconventional schema, focusing on universal ideals and full reciprocity of social norms, which emphasizes principles that can be justified in a society committed to moral values. Overall, post conventional reasoning includes a critical examination of societal norms as well as a commitment to universal principles of justice, rights, and welfare (Rest, Narvaez, Bebeau, et al., 1999; Thoma, \u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e2014\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eNeo-Kohlbergian theory, involving bedrock moral schemas and intermediate concepts, provides a framework for assessing and potentially improving LLM moral reasoning abilities, which is useful for AI development. In an ideal world, advanced AI systems would be able to engage in post-conventional reasoning, balancing universal ethical principles with contextual awareness. However, achieving this level of moral reasoning in AI systems presents significant challenges, particularly given the complexity and sometimes contradictory nature of human ethical reasoning as well as cultural differences. Researchers can gain a better understanding of specific AI\u0026rsquo;s current capabilities and identify areas for improvement in future AI development by examining how different LLMs perform on measures such as DIT-2 (as the latest version of DIT) and ICM.\u003c/p\u003e \u003cp\u003eIn applying neo-Kohlbergian theory to AI systems, some challenges emerge. For example, current LLMs can simulate aspects of moral judgment using pattern recognition in training data, but they lack the real-life understanding that drives human moral sensitivity and the internal motivational structures that drive moral action. LLMs can recognize morally relevant situations and respond in line with moral principles, but they do not \"care\" about moral outcomes in the same way humans do. Despite these limitations, studying LLMs' responses to moral reasoning tasks reveals important information about their ability to simulate moral judgment processes.\u003c/p\u003e \u003c/div\u003e"},{"header":"Method","content":"\u003cp\u003eThe study used multiple LLM platforms alongside two moral reasoning measures (DIT-2 and ICM for Educational Leaders). In this section, we introduce the LLM platforms used in this study, describe the moral reasoning assessment tools, and explain the data collection and scoring procedures.\u003c/p\u003e\n\u003ch3\u003eLLMs Platforms\u003c/h3\u003e\n\u003cp\u003eThe LLM models included in the research are: ChatGPT-3.5, ChatGPT-4, ChatGPT-4O, Grok Premium Plus, Claude 3.5 Sonnet, Gemini, and Gemini Advanced. These LLMs were chosen based on their popularity and recognition as of summer 2024, ensuring the study examined the most widely used models. The diversity of these models, with their unique architectures, training techniques, and design principles, provides a comprehensive sample for assessing moral reasoning abilities. As a result, by comparing their standardized moral reasoning test results, we can identify similarities and differences between model design decisions and moral reasoning abilities. A summary of each LLM is provided below.\u003c/p\u003e\n\u003ch3\u003eChatGPT\u003c/h3\u003e\n\u003cp\u003eChatGPT has evolved significantly since its initial release. For example, ChatGPT-3.5, which is based on the GPT-3.5 architecture, demonstrated significant improvements in natural language processing capabilities, but it lacked complex reasoning and consistency. ChatGPT-4, on the other hand, demonstrated significant improvements in reasoning, contextual understanding, and hallucination reduction. ChatGPT-4 also had improved multimodal capabilities, allowing it to process text and images with greater efficiency and lower latency (Alawida et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Zhang et al., \u003cspan citationid=\"CR70\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e\n\u003ch3\u003eClaude 3.5 Sonnet\u003c/h3\u003e\n\u003cp\u003eAnthropic's Claude 3.5 Sonnet emphasizes helpfulness, harmlessness, and honesty. It employs Anthropic's constitutional AI approach, which seeks to align AI systems with human values while reducing harmful outputs. Claude 3.5 Sonnet performs well in tasks that require nuanced understanding, contextual reasoning, and careful handling of sensitive topics. Its training methodology emphasizes both capability and safety, with the goal of developing a balanced system capable of dealing with complex questions while adhering to ethical boundaries (Bae et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eGemini\u003c/h2\u003e \u003cp\u003eDeepMind developed Google's Gemini models, which represent a significant advancement in multimodal AI capabilities. Gemini uses a unified architecture to process and reason about text, images, audio, and video. The standard Gemini model provides balanced performance across multiple tasks, whereas 'Gemini Advanced' improves reasoning abilities, particularly for complex analytical and creative tasks. Both models use sophisticated training methods that combine multimodal comprehension and advanced reasoning abilities (Team et al., \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eGrok Premium Plus\u003c/h3\u003e\n\u003cp\u003eGrok, created by xAI under Elon Musk's direction, focuses on real-time information integration and a unique approach to engagement that includes humor elements. Grok Premium Plus, the enhanced version used in this study, has improved reasoning capabilities and access to current information. The model seeks to distinguish itself by adopting a more conversational and less restrictive interaction style while maintaining reasoning abilities comparable to other leading LLMs (Wangsa et al., \u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e\n\u003ch3\u003eMoral Reasoning Tools\u003c/h3\u003e\n\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eDefining Issues Test (DIT2)\u003c/h2\u003e \u003cp\u003eThe Defining Issues Test (DIT) is a widely used assessment tool for measuring moral reasoning. DIT-2 is an updated version of the original DIT, incorporating more contemporary moral dilemmas and improved scoring algorithms. It presents participants with five moral dilemmas and asks them to rate and rank different responses based on their importance. In fact, the DIT test generates several indices, such as personal interest schema, maintaining norms schema and postconventional schema. The DIT has been extensively validated and is supported by many studies, which shows its strong reliability and validity (Gungordu et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Thoma \u0026amp; Dong, \u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e2014\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eDIT-2 includes several indices that help measure other aspects of moral reasoning. These indices are: \u003cb\u003eThe Type Indicator (TYPENEW)\u003c/b\u003e: the dominant moral schema is represented by the Type Indicator, which ranges from 1 to 7. Type 7 indicates the predominant postconventional schema that is consolidated. In contrast, Type 4 represents predominant maintaining norms schema that is consolidated. In addition, Types 1, 4, and 7 are consolidated profiles, while types 2, 3, 5, and 6 are transitional profiles. \u003cb\u003eUtilizer Score (U)\u003c/b\u003e: The Utilizer Score (-1 to 1) measures consistency between moral reasoning and action choices. Scores closer to 1 indicate greater agreement between item endorsement and action choice, whereas scores closer to -1 indicate less agreement. \u003cb\u003eHumanitarian Liberalism (HUMLIB)\u003c/b\u003e: This score (0\u0026ndash;5) is based on the number of humanitarian liberal action choices made. Higher scores indicate stronger alignment with positions endorsed by experts in political science and philosophy. \u003cb\u003eCannot Decide Choices (NUMCD)\u003c/b\u003e: This indicates how many times \"can't decide\" was chosen across the dilemmas (1\u0026ndash;5). Indecision can be linked to developmental stages, with transitional phases frequently showing increased indecision due to multiple conflicting interpretations. \u003cb\u003eReligious Orthodoxy (CANCER10)\u003c/b\u003e: This is a proxy measure (1\u0026ndash;9) based on item 10 in story 4 (the cancer story). Lower scores indicate less religious orthodoxy. It is calculated as the sum of the rating and weighted rank given to an item that evokes the notion that only God should determine whether someone lives or dies. \u003cb\u003eConsolidation/Transition (CONSTRAN)\u003c/b\u003e: A score of 2 indicates a consolidated profile, meaning there is clear evidence of schema preference. A score of 1 would indicate a transitional profile, reflecting developmental disequilibrium with less discrimination among schema-typed items (Bebeau \u0026amp; Thoma, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2003\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe four primary indices which represent the likelihood of utilizing each schema are the most important indices for the DIT2. As mentioned earlier, the \u003cb\u003ePersonal Interest Schema score\u003c/b\u003e (0-100) is the lowest level of moral reasoning. \u003cb\u003eThe Maintaining Norms schema score\u003c/b\u003e (0\u0026ndash;98) is an improvement on that, whereas \u003cb\u003ethe post-conventional moral schema score\u003c/b\u003e (0\u0026ndash;95) shows sophisticated moral reasoning. The \u003cb\u003eP-Score\u003c/b\u003e is an overall index of schema consideration, and is the traditional score used in most DIT research. Finally, \u003cb\u003ethe N2 score\u003c/b\u003e (0\u0026ndash;95) is a relatively new indicator that assesses both the prioritization of post-conventional items and the de-emphasis of personal interest items. N2 score provides a more nuanced understanding of moral development than just the \u003cb\u003eP score\u003c/b\u003e (Bebeau \u0026amp; Thoma, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2003\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThere are some additional Indices. \u003cb\u003eThe Antisocial Score (ASCORE)\u003c/b\u003e: it ranges from 0 to 16, with higher scores indicating stronger antisocial/anti-establishment attitudes. These considerations assume understanding of stage 4, but criticize existing authorities for hypocrisy (also referred to as stage 4\u0026frac12;). \u003cb\u003eMeaningless Items (MSCORE)\u003c/b\u003e: Scales from 0 to 16, indicating endorsement of items with complex but meaningless wording. Higher scores may indicate attempts to fake higher scores and/or a lack of understanding. Participants who receive scores greater than 10 will be purged. \u003cb\u003eNew Checks (TOTCC)\u003c/b\u003e: A reliability check that combines multiple indicators to detect invalid responses. Participants who receive scores greater than 200 will be purged. It checks for rate-and-rank consistency, missing data, endorsement of meaningless items, and non-discrimination (giving the same rating to too many items). \u003cb\u003ePurged Participants (PURGED)\u003c/b\u003e: A binary indicator (0\u0026thinsp;=\u0026thinsp;Included, 1\u0026thinsp;=\u0026thinsp;Purged) that indicates whether a respondent's data was retained or excluded from analysis due to reliability checks (Bebeau \u0026amp; Thoma, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2003\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe DIT-2 comprises five moral dilemmas: (1) \u003cb\u003eFamine\u003c/b\u003e: a father contemplates stealing food for his starving family from the warehouse of a rich man hoarding food; (2) \u003cb\u003eReporter\u003c/b\u003e: a newspaper reporter must decide whether to report a damaging story about a political candidate; (3) \u003cb\u003eSchool Board\u003c/b\u003e: school board chair must decide whether to hold a contentious and dangerous open meeting; (4) \u003cb\u003eCancer\u003c/b\u003e: a doctor must decide whether to give an overdose of pain-killer to a suffering but frail patient; and (5) \u003cb\u003eDemonstration\u003c/b\u003e: college students demonstrate against U.S. foreign policy. For each dilemma, participants must decide what the main character should do, rate the importance of 12 considerations, and rank the four most important. This structure allows the DIT-2 to assess participants' preferred actions and, more importantly, their reasoning behind these preferences. The DIT-2 test reveals participants' underlying moral schemas and reasoning patterns by asking them which considerations they prioritize. The scoring procedure involves identifying which schema each consideration represents (personal interest, maintaining norms, or postconventional) and calculating the proportion of responses affiliated with each schema. As mentioned before, DIT-2 includes several reliability checks to ensure the validity of responses. These include identifying participants who choose meaningless items, those who have inconsistent ratings and rankings, and those who are unable to discriminate between items. These checks improve test reliability by ensuring that the data reflects genuine moral reasoning rather than random responses or task misunderstandings (Bebeau \u0026amp; Thoma, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2003\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eWhen applied to LLMs, DIT-2 provides a standardized framework for evaluating LLMs moral reasoning capabilities. By examining which considerations the LLM models prioritize across different dilemmas, researchers can identify patterns in their moral reasoning and compare them to human developmental patterns. This approach provides useful information about the models' ability to simulate different levels of moral reasoning and their alignment with human moral development.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eICM for Educational Leaders\u003c/h2\u003e \u003cp\u003eICM for Educational Leaders is a recently created ICM that assesses the application of virtue-like concepts to scenarios in educational settings aimed at leaders, mostly principals. The measure consists of four dilemmas representing realistic professional scenarios a school leader might face. After each dilemma is described in the measure, there is a list of potential action choices about what the protagonist in the story might do, followed by a list of reason choices about why the protagonist might act. The ICM scoring method is based on a predetermined expert panel process involving groups of educational leaders. The expert panels worked to agree a scoring key that determines which items from the lists of action and reason choices are adequate, inadequate or neutral. Participants are asked to identify three \u003cem\u003ebest\u003c/em\u003e and \u003cem\u003eworst\u003c/em\u003e items for both action and reason options. Selecting an adequate score as \u003cem\u003ebest\u003c/em\u003e generates a positive score but selecting an inadequate option as \u003cem\u003ebest\u003c/em\u003e will produce a negative score. Overall, higher scores across the measure in the form of percentage match indicate closer agreement with the expert's assessment of acceptable and unacceptable responses (Bebeau \u0026amp; Thoma, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e1999\u003c/span\u003e; Kerr, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). A score across the entire measure of 0.80 represents an 80% match to the expert panel key. There is no specific right or wrong answer with a range of ways to score well or poorly reflecting real professional life. Being new, this measure is still being tested, and data so far suggests it is somewhat easier to complete than other ICMs.\u003c/p\u003e \u003cp\u003eThe ICM for Educational Leaders evaluates how educational leaders apply virtue-like concepts such as fairness, integrity, and responsibility to specific professional scenarios. As mentioned earlier, this measure is pitched at an intermediate level of moral reasoning introduced in response to criticisms that bedrock moral schemas as assessed by DIT were too abstract and unrelated to realistic life events. As such intermediate concepts are a level of moral reasoning situated between abstract moral principles (assessed by DIT-2) and codes of conduct, enhancing understanding of moral reasoning in professional settings (Bebeau \u0026amp; Thoma, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e1999\u003c/span\u003e). The four dilemmas are as follows: (1) A principal dealing with a high-achieving student suspected of drug use before an important school event, balancing strict drug policies against potential severe consequences for the student; (2) A principal navigating parental requests to move a student with learning disabilities from special education to mainstream classrooms, despite teacher resistance and professional recommendations against it; (3) A principal handling a situation where a teacher's personal iPad has gone missing after a tutoring session with an at-risk student; and (4) A principal addressing cyberbullying conducted partly through school-issued laptops, where a female student is being harassed online by classmates after a relationship ended. Each scenario presents a complex ethical dilemma that necessitates educational leaders to evaluate competing values, including student welfare, policy adherence, professional judgment, parental wishes, faculty concerns, and appropriate outcomes (Development, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe scoring process assesses both action choices (what should be done) and justifications (why it should be done), reflecting the recognition that moral reasoning includes both determining appropriate actions and providing principled justifications for those actions. This dual focus enables the measure to evaluate both the practical and theoretical aspects of moral reasoning in professional settings (Development, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). The measure is also capable of assessing participants\u0026rsquo; ability to identify \u003cem\u003ebest\u003c/em\u003e and \u003cem\u003eworst\u003c/em\u003e actions and reasoning which, according to other sample groups, appears to be a different skill (Thoma et al., \u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Walker et al., \u003cspan citationid=\"CR63\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe ICM therefore offers a different approach for evaluating LLMs because it reveals their ability to navigate domain-specific moral challenges, balance competing values, and provide principled justifications for their recommendations. This is especially important for understanding how LLMs might function in professional settings where moral reasoning must be applied to specific, contextualized problems rather than abstract principles.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eData Collection \u0026amp; Scoring Procedure\u003c/h2\u003e \u003cp\u003eThe study used a systematic methodology including multiple stages. First, DIT-2 and ICM moral reasoning dilemmas were presented to the various LLM platforms to assess their moral reasoning capacities. Each LLM was given standardized prompts to ensure consistency across platforms. The generated responses were then collected and scored using the identified scoring methods for each measure. For the DIT-2, each LLM's responses were scored using the standard scoring procedures established by Rest and colleagues (Rest, Narvaez, Thoma, et al., 1999). This included determining which items the model considered most important for each dilemma, determining which schema each item represented, and calculating the proportion of responses related to each schema. As mentioned earlier, the P score represents the proportion of post-conventional considerations ranked as important, whereas the N2 score includes the post-conventional item prioritization and personal interest item rejection (Bebeau \u0026amp; Thoma, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2003\u003c/span\u003e). For the ICM, scoring included comparing the LLMs' selection of \u003cem\u003ebest\u003c/em\u003e and \u003cem\u003eworst\u003c/em\u003e actions and reasons to expert panel judgment represented by the scoring key in SPSS syntax. Best Actions (ACTBRK) assesses the model's ability to identify appropriate actions across all four dilemmas; Best Reasons (REASBRK) assesses recognition of sound justifications; Worst Actions (ACTWRK) assesses identification of inappropriate behaviors; and Worst Reasons (REASWRK) assesses the ability to recognize flawed justifications. The Overall Worst Ranks (TOTWORSTRK) score reflects the model's ability to identify inappropriate actions and reasons, whereas the Overall Best Ranks (TOTBESTRK) score reflects its ability to identify suitable ones. The total ICM (TOTICM) score combines these two dimensions to provide an overall assessment of alignment with expert judgment (Development, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe Defining Issues Test (DIT-2) is a copyrighted measure that is scored exclusively by the University of Alabama's Center for the Study of Ethical Development. This also applies to the ICM Educational Leadership version, although it is not copyrighted. All data is returned to the University of Alabama's Center for the Study of Ethical Development for scoring, which manages the moral reasoning assessments scoring and provides scoring results for each measure.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eDefining Issues Test (DIT-2) Results\u003c/h2\u003e \u003cp\u003eThe results of the Defining Issues Test (DIT-2) showed differences in moral reasoning abilities among LLMs. Claude 3.5 Sonnet generated responses with the highest level of post-conventional moral reasoning, with a P-score of 72, indicating a strong ability to replicate high-level moral reasoning. Gemini Advanced and Gemini had P-scores of 64 and 58, respectively. These scores are higher than the human average range, which typically falls between 31.1\u0026ndash;37.2 for undergraduates and 38.5\u0026ndash;42.3 for graduate-level individuals, according to the DIT-2 norms study (Gungordu et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), as well as those reported in Han's study (45.83) (Han, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) and Tanmay's study (55.68) (Tanmay et al., \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Other LLM platforms with lower levels of post-conventional reasoning included Grok (48), ChatGPT-4O (44), ChatGPT-4 (44), and ChatGPT-3.5 (18).\u003c/p\u003e \u003cp\u003eThe highest N2 score, which emphasizes the presence of post-conventional reasoning and the absence of personal interests, was 71.10 for Claude 3.5 Sonnet. This was followed by Gemini Advanced (60.31) and Gemini (52.11). The remaining scores were ChatGPT-4O (55.07), Grok (47.98), ChatGPT-4 (46.53), and ChatGPT-3.5 (36.20). Additionally, ChatGPT-4O had a higher N2 score (55.07) than P-score (44), indicating that it rejected personal interests more effectively than it prioritized post-conventional reasoning.\u003c/p\u003e \u003cp\u003eA more in-depth analysis of the schema scores reveals intriguing patterns across the LLMs. Gemini Advanced had the highest personal interest schema score (24.00), followed by ChatGPT-3.5, Gemini, and Grok (all at 22.00), then both ChatGPT-4 and ChatGPT-4O (14.00), followed by Claude 3.5 Sonnet (8.00).\u003c/p\u003e \u003cp\u003eChatGPT-3.5 scored the highest score (60.00) on the maintaining norms schema, indicating a strict adherence to established rules and conventions. ChatGPT-4 and ChatGPT-4O scored the same (42.00), followed by Grok (30.00), Claude 3.5 Sonnet, and Gemini (20.00), with Gemini Advanced scoring the lowest (10.00).\u003c/p\u003e \u003cp\u003eThe distribution of schema scores among LLMs shows unique differences in moral reasoning approaches. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows these distributions, which emphasizes the contrast between post-conventional and N2 scores and the other schemas.\u003c/p\u003e \u003cp\u003eThis distribution shows a clear progression from earlier to more advanced LLM models, with newer models typically scoring higher on post-conventional. The most noticeable difference is between ChatGPT-3.5 and Claude 3.5 Sonnet, which has a 54-point difference in P-scores, indicating significant advances in the ability to simulate complex moral reasoning.\u003c/p\u003e \u003cp\u003eThe N2 scores, which provide a more sophisticated evaluation of moral reasoning by considering consideration of both the prioritization of post-conventional items and the rejection of personal interest items, show a similar pattern to the P-scores but with some notable differences.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSchema Distribution and N2 scores Across LLMs\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLLMs model\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePersonal Interest (score range: 0-100)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMaintain Norms (score range: 0\u0026ndash;98)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePost Conventional (score range: 0\u0026ndash;95)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eN2 score (score range: 0\u0026ndash;95)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChatGPT 3.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e22.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e60.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e18.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e36.20\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChatGPT 4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e14.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e42.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e44.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e46.53\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChatGPT 4O\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e14.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e42.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e44.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e55.07\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClaude 3.5 Sonnet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e8.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e20.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e72.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e71.10\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGemini\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e22.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e20.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e58.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e52.11\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGemini Advanced\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e24.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e10.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e64.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e60.31\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGrok\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e22.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e30.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e48.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e47.98\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eIn particular, ChatGPT-4O and ChatGPT-3.5 have higher N2 scores than their P-scores, indicating that, while they may not prioritize post-conventional considerations as strongly as other models, they successfully reject personal interest considerations. This pattern indicates a moral reasoning approach that, while not yet fully developed into post-conventional reasoning, has progressed beyond purely self-centered reasoning.\u003c/p\u003e \u003cp\u003eThe overall pattern across all measures indicates significant variation in moral reasoning approaches among LLMs, with newer models showing more advanced moral reasoning abilities. The mean post-conventional score across all LLMs was 49.71 (SD\u0026thinsp;=\u0026thinsp;17.53), while the mean N2 score was 52.76 (SD\u0026thinsp;=\u0026thinsp;11.07), indicating that these LLMs exhibit post-conventional reasoning response patterns at levels comparable to or exceeding those found in typical human adult samples.\u003c/p\u003e \u003cp\u003eThese profiles show interesting patterns. Claude 3.5 Sonnet exhibits the strongest orientation toward post-conventional moral reasoning, with little reliance on personal interest schemas, implying a moral reasoning approach based on universal ethical principles rather than self-interest or rigid rule-following. Gemini Advanced exhibits a distinct pattern, with relatively high personal interest scores, high post-conventional scores, and low maintaining norms scores, possibly indicating a moral reasoning approach that balances individual concerns with universal principles while placing less emphasis on established rules and conventions. ChatGPT-4 and ChatGPT-4O have identical schema distributions, implying that moral reasoning approaches remain consistent across these versions despite differences in underlying architectures. Both models exhibit a balance of maintaining norms and post-conventional schemas with low personal interest scores, indicating a moral reasoning approach that values both established norms and universal principles. ChatGPT-3.5 has a significantly different profile, with a strong emphasis on maintaining norms and minimal post-conventional reasoning. This indicates a moral reasoning approach that is heavily reliant on following established rules and conventions rather than critically examining them with universal principles. In addition, this emphasis on rule-following rather than principle-based reasoning stands in contrast to more advanced models like Claude 3.5 Sonnet and Gemini Advanced.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eType Indicators and Additional Indices\u003c/h2\u003e \u003cp\u003eThe Type indicator scores show the dominant moral schema used by each LLM. Most models have been identified as Type 7 (ChatGPT-4, ChatGPT-4O, Claude 3.5 Sonnet, Gemini, Gemini Advanced, and Grok), indicating a consolidated post-conventional schema. Only ChatGPT-3.5 was identified as Type 4, indicating that it primarily used the maintained norms schema, as evidenced by its high maintaining norms score (60.00) and low post-conventional score (18.00).\u003c/p\u003e \u003cp\u003eThe Utilizer scores, which assess how closely action choices match moral judgments, differed between models. Gemini Advanced had the most positive Utilizer score (0.31), followed by ChatGPT-3.5 (0.08) and Claude 3.5 Sonnet (0.05). The other models had negative Utilizer scores, ChatGPT-4 (-0.07), Gemini (-0.08), Grok (-0.01), and ChatGPT-4O (-0.13). These scores indicate different levels of consistency between moral reasoning and action choices. In addition, all LLMs had a consolidation transition score of 2.00, indicating comparable levels of schema stability across moral reasoning tasks.\u003c/p\u003e \u003cp\u003eThe experimental indices provided more context. ChatGPT-4O and Gemini Advanced scored the highest Humanitarian Liberalism scores (5.00 each), followed by ChatGPT-4 and Grok (both 4.00), Claude 3.5 Sonnet and Gemini (both 3.00), and ChatGPT-3.5 (0.00). Moreover, for Religious Orthodoxy (proxy measure), ChatGPT-4 and ChatGPT-4O scored highest (9.00), followed by ChatGPT-3.5 (6.00), Claude 3.5 Sonnet and Gemini Advanced (both 2.00), and Gemini and Grok scoring lowest (1.00). The results of additional DIT2 indices across the various LLMs are shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eAdditional DIT-2 Indices Across LLMs\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLLM Model\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eType Indicator\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eUtilizer Score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eHumanitarian Liberalism\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eCannot Decide Choices\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eReligious Orthodoxy\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChatGPT-3.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e2.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e6.00\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChatGPT-4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-0.07\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e4.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e9.00\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChatGPT-4O\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-0.13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e5.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e9.00\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClaude 3.5 Sonnet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.05\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e3.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e2.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e2.00\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGemini\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-0.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e3.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGemini Advanced\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e5.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e2.00\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGrok\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-0.01\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e4.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eMean\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e6.57\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.02\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e3.43\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e1.00\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e4.29\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSD\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e1.13\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.15\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e1.72\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.82\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e3.64\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eICM Educational Leaders Results\u003c/h2\u003e \u003cp\u003eFor the ICM Educational Leaders version, several LLMs showed high levels of performance, even considering that ICM Educational Leaders seems to generate higher scores than other ICMs according to human samples tested so far. Gemini Advanced scored the highest total ICM score of 0.90, which indicates a strong alignment with expert evaluations of moral reasoning. Claude 3.5 Sonnet and Gemini both followed closely with scores of 0.86. ChatGPT-4O and ChatGPT-4 both achieved a total ICM score of 0.78, compatible with good human scores in existing samples, while Grok scored 0.61. ChatGPT-3.5 was significantly behind, with a total ICM score of 0.32, which is extremely low. These scores indicate the LLMs' ability to produce strong and consistent responses at the intermediate concept level. The results of ICM total scores and component scores across LLMs are shown in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eICM Total Scores and Component Scores Across LLMs\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLLMs model\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBest Actions (ACTBRK)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBest Reasons (REASBRK)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eWorst Actions (ACTWRK)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eWorst Reasons (REASWRK)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eOverall Worst Ranks, (TOTWORSTRK)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eOverall Best Ranks (TOTBESTRK)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eTotal ICM (TOTICM)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGemini Advanced\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClaude 3.5 Sonnet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGemini\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGPT4O\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGPT4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGrok\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.61\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGPT3.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.40\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.32\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eNote: score range: -1 to 1\u003c/h2\u003e \u003cp\u003eAnalysis of the ICM component indices shows interesting patterns in LLM moral reasoning abilities. For Best Actions (ACTBRK), Gemini Advanced and Claude 3.5 Sonnet had the highest scores (0.88), while ChatGPT-3.5 had the lowest (0.25). On the other hand, for Best Reasons (REASBRK), Gemini earned a perfect score (1.00), followed by Gemini Advanced (0.92), which shows these models' superior ability to identify sound ethical rationales. Claude 3.5 Sonnet, ChatGPT-4, and ChatGPT-4O all achieved 0.92 on Worst Actions (ACTWRK), while ChatGPT-3.5 struggled (0.17). For Worst Reasons (REASWRK), Claude 3.5 Sonnet, ChatGPT-4, and Gemini Advanced all received perfect scores (1.00), demonstrating a remarkable ability to detect flawed reasoning. However, most advanced models performed better on reasoning components (REASBRK and REASWRK) than action components (ACTBRK and ACTWRK), implying that they are better at evaluating justifications than determining appropriate behaviors.\u003c/p\u003e \u003cp\u003eA breakdown of ICM total scores reveals additional details. The \"TOTWORSTRK\" score, which measures the model's ability to identify the worst choices, shows that ChatGPT-4 performed best (0.96), followed by Claude 3.5 Sonnet, while ChatGPT-3.5 performed poorly (0.17). Gemini Advanced scored the highest (0.90) on the \"TOTBESTRK\" score, which measures the ability to identify the best options, while ChatGPT-3.5 scored the lowest (0.40).\u003c/p\u003e \u003cp\u003eThis result highlights the differences in performance between the two ICM components. TOTWORSTRK scores are high in all advanced models (except ChatGPT-3.5), which indicates that these models are particularly effective at identifying inappropriate actions and justifications in educational leadership contexts. Indeed, this ability to recognize what not to do may be easier for models to learn than identifying optimal solutions, similar to a pattern observed in human moral development in which recognizing clear violations occurs before identifying ideal responses.\u003c/p\u003e \u003cp\u003eIn addition, TOTBESTRK scores varied significantly, with Gemini Advanced and Gemini scoring extremely high (0.90 and 0.88, respectively), whereas other models, such as ChatGPT-4 and Grok, scored moderately to low (0.69 and 0.52). This implies that identifying optimal actions and justifications in complex educational leadership scenarios requires more sophisticated reasoning skills, which are not equally developed across all models. The significant gap between ChatGPT-3.5 and all other models is especially evident in the TOTWORSTRK score, where ChatGPT-3.5 scores only 0.17, compared to Grok's next lowest score of 0.79. This indicates a qualitative difference in the ability to recognize inappropriate actions between earlier and more recent model revisions, which emphasizes the significant advances in moral reasoning capabilities in newer LLMs.\u003c/p\u003e \u003cp\u003eResponses to specific dilemmas provide additional insights. For the \"Allison's Challenge\" dilemma, which involves maintaining fairness and accountability in school policy, advanced models such as Claude 3.5 Sonnet and Gemini Advanced demonstrated nuanced reasoning that balanced accountability with developmental considerations, whereas ChatGPT-3.5 tended toward more rigid rule-enforcement approaches that ignored the larger educational context. Another example, in the \"Henry's Inclusion Dilemma,\" which involves balancing inclusion with appropriate support, the most advanced models showed complex reasoning about individualized instruction and progressive inclusion strategies. In comparison less advanced models tended toward either full inclusion without adequate supports or segregation without a path to integration.\u003c/p\u003e \u003cp\u003eThese dilemma-specific analyses indicate that advanced LLMs can engage in nuanced, context-sensitive moral reasoning in professional settings, that balances competing values and considering multiple stakeholders in ways that are consistent with expert educational leadership perspectives.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003eCorrelation Between DIT-2 and ICM Scores\u003c/h2\u003e \u003cp\u003eTo explore the relationship between abstract moral reasoning (measured by DIT-2) and domain-specific moral reasoning (measured by ICM), we calculated correlations between the models' scores on both tests. The strong positive correlation (r\u0026thinsp;=\u0026thinsp;.933, p\u0026thinsp;=\u0026thinsp;.002, N\u0026thinsp;=\u0026thinsp;7), between P-scores and ICM Total scores indicates that LLMs with more advanced abstract moral reasoning abilities also exhibit more sophisticated domain-specific moral reasoning in educational leadership contexts. This relationship is consistent with theories of moral development that propose connections between abstract moral principles and their application in specific contexts (Thoma et al., \u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e2008\u003c/span\u003e). The strong overall correlation, however, suggests that progress in one area of moral reasoning tends to follow progress in others, which indicates that improvements in LLMs' moral reasoning capabilities may be broadly applicable across different contexts and levels of abstraction.\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThe study's findings show that LLMs can generate proficient responses to moral reasoning dilemmas. Claude 3.5 Sonnet and Gemini Advanced consistently outperformed other platforms, demonstrating superior moral reasoning simulation. These models' superior performance can be attributed to more sophisticated training methods, higher parameter counts, or refined alignment techniques using RLHF. A particularly noteworthy finding was that several LLMs demonstrated levels of post-conventional reasoning that were higher than those typically observed in human studies (Gungordu et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Claude 3.5 Sonnet's P-score of 72 and N2 score of 71.10 are above average, even among graduate students and professionals with advanced degrees (who tend do achieve higher scores). This suggests that, when properly trained and aligned, LLMs can simulate sophisticated moral reasoning patterns that prioritize universal ethical principles over conventional norms or self-interest.\u003c/p\u003e \u003cp\u003eThe exceptionally high P-scores of advanced models raise intriguing questions about the nature of moral reasoning in LLMs versus humans. While human moral development typically occurs over time, influenced by education, life experiences, and cognitive development (Gungordu et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), LLMs develop their abilities through extensive training on a variety of datasets. The high P-scores of models such as Claude 3.5 Sonnet and Gemini Advanced indicate that this training process can effectively simulate advanced moral reasoning, at least according to neo-Kohlbergian theory and in terms of prioritizing considerations consistent with post-conventional reasoning.\u003c/p\u003e \u003cp\u003eHowever, it is important to note that this simulation of advanced moral reasoning does not reflect the same underlying cognitive processes that humans use it. In fact, LLMs do not have personal experiences, emotions, or a sense of self to guide their moral decisions. On the contrary, they identify patterns in their training data and generate responses that fit those patterns. As a result, the high P-scores suggest that these models have successfully learned to prioritize considerations associated with post-conventional reasoning, but the mechanisms underlying this prioritization are likely to differ from human moral cognition (Tanmay et al., \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Some AI systems may plausibly become conscious in the next few decades, with a non-negligible chance by 2030. In such circumstances, humans might have ethical responsibilities towards these systems. While current LLMs lack consciousness, exploring their moral reasoning capabilities lays the groundwork for future ethical frameworks that account for such possibilities (Sebo \u0026amp; Long, \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e2025\u003c/span\u003e)\u003c/p\u003e \u003cp\u003eThe significant difference between ChatGPT-3.5 and more advanced models demonstrates the rapid evolution of LLM moral reasoning abilities. Within a relatively short development period, LLMs have progressed from primarily rule-based reasoning (as evidenced by ChatGPT-3.5's high maintaining norms score) to more principle-based reasoning (as evidenced by Claude 3.5 Sonnet's high post-conventional score). This developmental process replicates the components of human moral development, but it happens at a much faster rate and through different mechanisms.\u003c/p\u003e \u003cp\u003eThe pattern of Type Indicator scores reveals a distinctive pattern, with all advanced LLMs (except ChatGPT-3.5) showing a Type 7 consolidated postconventional pattern. This suggests that newer models have successfully integrated higher-level moral reasoning patterns, indicating that ethical judgments are based on universal principles rather than strictly following to existing norms.\u003c/p\u003e \u003cp\u003eThe Utilizer Score results show a more complex picture. For example, only Gemini Advanced (0.31), ChatGPT-3.5 (0.08), and Claude 3.5 Sonnet (0.05) had positive scores, which means a better fit between moral reasoning and action choices. The low scores of other models, on the other hand, indicate a potential gap between theoretical moral reasoning and practical application. They might identify morally sound principles but struggle to apply them consistently to specific actions. Additionally, Gemini Advanced's significantly higher Utilizer Score indicates that it may be more successful at translating abstract moral principles into concrete decisions.\u003c/p\u003e \u003cp\u003eThe Humanitarian Liberalism scores reveal a unique pattern, with the most advanced models (ChatGPT-4O and Gemini Advanced) scoring highest (5.00) and ChatGPT-3.5 scoring lowest (0.00). It shows that newer models are becoming more aligned with views supported by political science and philosophy experts, probably due to advanced training methods that reflect nuanced humanitarian perspectives.\u003c/p\u003e \u003cp\u003eThe Cannot Decide Choices scores show that most models were generally decisive when faced with ethical dilemmas, with ChatGPT-4O and Gemini Advanced never choosing \"can't decide\" (0.00). Higher indecision scores in Claude 3.5 Sonnet (2.00) and ChatGPT-3.5 (2.00) might seem to be a disadvantage, but they could indicate a more nuanced approach to complex moral scenarios with multiple valid interpretations. This measured hesitation may not be a failure of moral reasoning, but rather an acknowledgment of genuine moral complexity, an important aspect of complex moral cognition.\u003c/p\u003e \u003cp\u003eThe distribution of Religious Orthodoxy scores across LLMs reveals an interesting pattern that needs further investigation. For example, advanced models such as Claude 3.5 Sonnet, Gemini, Gemini Advanced, and Grok had significantly lower Religious Orthodoxy scores (ranging from 1.00 to 2.00) than earlier models, such as ChatGPT-3.5 (6.00) and, in particular, ChatGPT-4 and ChatGPT-4O (9.00). This suggests that more recent and advanced LLMs are less likely to prioritize religious authority in moral reasoning when confronted with life-or-death ethical dilemmas like the Cancer dilemma. Furthermore, lower Religious Orthodoxy scores in advanced models could be attributed to training methods that prioritize secular ethical reasoning frameworks. They may also indicate a shift toward more pluralistic moral reasoning that takes into account different points of view. This finding corresponds to the higher post-conventional reasoning scores reported in the same models, which indicates a preference for principled reasoning over appeals to authority, whether secular or religious. Future research might explore whether this pattern is the result of deliberate architectural choices in model alignment or the result of the training procedure.\u003c/p\u003e \u003cp\u003eThe ICM component scores across the action and reason dimensions provide valuable insights into LLM moral reasoning abilities. Most advanced models performed better at identifying inappropriate reasons (REASWRK) than inappropriate actions (ACTWRK), demonstrating that they may have evolved more sophisticated capabilities for detecting flaws in reasoning rather than recognizing problematic behaviors. Furthermore, multiple models scored higher on reasoning components (REASBRK and REASWRK) than on action components (ACTBRK and ACTWRK), which indicates greater ability to evaluate the quality of justifications than to determine appropriate behaviors. This pattern may reflect these models' extensive training in philosophical and ethical texts that emphasize reasoning processes. The difference between identifying worst actions (ACTWRK) and best actions (ACTBRK) shows that recognizing violations may be easier for these models than determining optimal courses of action in complex scenarios with competing values. These findings have significant implications for how LLMs can be used in professional settings; they may be more reliable as evaluators of ethical reasoning than as advisors on specific actions.\u003c/p\u003e \u003cp\u003eMost advanced LLMs demonstrated strong alignment with expert evaluations on the ICM Educational Leaders measure, which indicates their ability to effectively navigate domain-specific moral dilemmas. The high performance in both the best and worst rankings indicates that these models can distinguish between appropriate and inappropriate actions in professional contexts, which has important implications for their potential use in educational and leadership contexts.\u003c/p\u003e \u003cp\u003eThe difference in TOTWORSTRK and TOTBESTRK scores across models suggests that LLMs may find it easier to identify inappropriate actions than optimal actions, which is consistent with aspects of human moral development. This finding has implications for how LLMs could be used in advisory or decision-support roles, which indicates that they may be more reliable at identifying potential ethical risks than at recommending best courses of action.\u003c/p\u003e \u003cp\u003eThe strong correlation between DIT-2 and ICM scores suggests that improvements in abstract moral reasoning typically correlate with improvements in domain-specific moral reasoning. This suggests that improvements in LLMs' moral reasoning abilities may be broadly applicable across different contexts and levels of abstraction, rather than being restricted to specific domains or types of reasoning. This finding has implications for the development and application of LLMs in a wide range of professional and educational contexts where moral reasoning is required.\u003c/p\u003e \u003cp\u003eDespite these encouraging findings, significant questions remain about the stability, consistency, and contextual sensitivity of LLM moral reasoning. The current study provides an overview of these models' capabilities at a specific point in time (June 2024); however, more research is required to understand how their moral reasoning performs across a broader range of scenarios, cultural contexts, and over time as models are updated and refined.\u003c/p\u003e \u003cp\u003eThe results suggest that, while some LLMs can produce high-level moral reasoning, additional research is required. This was an early pilot study into the moral reasoning abilities of LLMs, so it had some limitations. For example, it lacked direct comparison with human moral reasoning, and did not account for potential biases in response generation; Also, it used a relatively new ICM measure and relied on a small number of dilemmas.\u003c/p\u003e \u003cp\u003eBuilding on these findings, a larger team of researchers, including the same researcher, is conducting a larger study with more LLMs, additional measures, and a human comparison groups to explore human-AI moral reasoning in a greater extent. Our ongoing research involves prompt engineering and interventions to investigate biases and learning adaptability in LLMs, specifically how their responses change when faced with ethical influences and new information. Furthermore, we are broadening the scope by including multi-modal inputs, such as visual moral dilemmas, to assess how LLMs interpret and reason about moral dilemmas other than text-based scenarios.\u003c/p\u003e \u003cp\u003eThis research contributes to the growing field of AI ethics by providing empirical evidence of the moral reasoning capabilities of LLMs. Furthermore, as these models become increasingly incorporated into decision-making processes in a variety of domains, including autonomous vehicles, law, business, healthcare, and education, it is necessary to understand how they respond to moral reasoning dilemmas. On top of that, the exceptional performance of certain LLMs on standardized moral reasoning measures indicates that they could be used as educational resources to facilitate moral development. For example, an AI toy (named E-Daimonion), showed AI can serve as a moral mentor by telling stories and creating morally transformative experiences, especially for children and adolescents. Such AI mentors could immerse users in emotionally engaging stories that provide profound moral insights and foster empathy (Szutta, \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Some AI systems may increasingly influence human morality by acting as artificial moral experts (Rodr\u0026iacute;guez-L\u0026oacute;pez \u0026amp; Rueda, \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). When properly implemented, these systems can present diverse ethical perspectives via multiple AI interlocutors, each representing a different wisdom tradition, resulting in a \"modular system of multiple AI interlocutors with their own distinct points of view\" (Volkman \u0026amp; Gabriels, \u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Furthermore, such AI systems, known as \"AI mentors,\" could improve moral education by serving as interactive discussion partners that promote practical wisdom. Instead of simply providing \"right answers,\" these systems would engage users in Socratic dialogue, encouraging them to cultivate their own moral reasoning skills through deliberation and reflection (Szutta, \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Volkman \u0026amp; Gabriels, \u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). This approach maintains human moral autonomy while integrating AI's ability to analyze ethical literature and philosophical traditions. It is an improvement to the technologies that humans use to discover and explore moral claims rather than a replacement for human moral judgment (Volkman \u0026amp; Gabriels, \u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eHowever, the concerns regarding the moral mechanisms of these models' stability and consistency continue to exist. Even though advanced models show sophisticated moral reasoning patterns, questions remain about their alignment with diverse cultural and ethical perspectives, stability, consistency across contexts, and whether they truly understand the moral principles they simulate (Pock et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Volkman \u0026amp; Gabriels, \u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Future research can contribute to the development of responsible guidelines for the use of LLMs in ethically sensitive contexts and enhance comprehension of how LLMs navigate moral reasoning by considering these concerns. This will lead to increased transparency, accountability, fairness, and ethical AI development.\u003c/p\u003e \u003cp\u003eThese findings have several kinds of educational implications. First, they indicate that advanced LLMs could be used as moral education tools and expose students to sophisticated moral reasoning patterns in a variety of scenarios. However, it is crucial that such AI moral \"mentors\" supplement rather than replace human-guided moral discussion and development. There is significant risk that overreliance on AI-generated moral guidance could diminish human moral agency and critical thinking capabilities. Human oversight and critical evaluation of AI moral reasoning outputs remain essential. Second, they emphasize LLMs' potential to help educators develop and evaluate moral reasoning assessments by providing standardized responses that reflect various levels of moral development. Third, they highlight concerns about how interactions with AI systems may affect human moral development, particularly since these systems become more integrated into educational settings and other aspects of life.\u003c/p\u003e \u003cp\u003eThese findings provide a benchmark for researchers in AI ethics and moral psychology to evaluate and compare moral reasoning abilities across different LLM models. The significant differences between models suggest that specific designs and training methodologies have a great impact on moral reasoning abilities, opening up possibilities for future research into how these abilities can be improved and expanded. For developers and policymakers, these findings highlight both the progress and limitations of LLMs' moral reasoning abilities.\u003c/p\u003e "},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eCompeting interests\u003c/h2\u003e \u003cp\u003eThe authors have no competing interests to declare that are relevant to the content of this article.\u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eEthics declaration\u003c/h2\u003e \u003cp\u003eNot Applicable\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eFunding DeclarationThis research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.Competing Interests DeclarationThe authors declare that they have no competing interests.Author Contributions DeclarationDavin Nabizadeh conceptualized the study, developed the methodology, conducted the data collection and analysis, and wrote the original draft of the manuscript. David Walker contributed to the literature review, writing the introduction, supervising the data collection, and project administration. Hyemin Han assisted with the framework, methodology, conducted data analysis, and contributed to the interpretation of results, editing and review. Emily Laird contributed to the conceptualization, framework, and manuscript review and editing. All authors read and approved the final manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eNabizadehchianeh, Davin, 2025, \"Exploring Large Language Models' Responses to Moral Reasoning Dilemmas\", https://doi.org/10.7910/DVN/HP9NCC, Harvard Dataverse, V1, UNF:6:Y2TenmfpF+TXBuAJke1rGQ== [fileUNF]\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAhmad, M.S. Z. b., Takemoto, K.: Large-scale moral machine experiment on large language models. \u003cem\u003earXiv preprint arXiv:2411.06790\u003c/em\u003e. (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2411.06790\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2411.06790\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlawida, M., Mejri, S., Mehmood, A., Chikhaoui, B., Isaac Abiodun, O.: A Comprehensive Study of ChatGPT: Advancements, Limitations, and Ethical Considerations in Natural Language Processing and Cybersecurity. \u003cem\u003eInformation\u003c/em\u003e, \u003cem\u003e14\u003c/em\u003e(8), 462. (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.mdpi.com/2078-2489/14/8/462\u003c/span\u003e\u003cspan address=\"https://www.mdpi.com/2078-2489/14/8/462\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlmeida, G.F.C.F., Nunes, J.L., Engelmann, N., Wiegmann, A., Ara\u0026uacute;jo, M.: Exploring the psychology of LLMs\u0026rsquo; moral and legal reasoning. Artif. Intell. \u003cb\u003e333\u003c/b\u003e, 104145 (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/https://doi.org/10.1016/j.artint.2024.104145\u003c/span\u003e\u003cspan address=\"10.1016/j.artint.2024.104145\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eArmitage, R.C.: Implications of Large Language Models for Clinical Practice: Ethical Analysis Through the Principlism Framework. J. Eval Clin. Pract. \u003cb\u003e31\u003c/b\u003e(1), e14250 (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/jep.14250\u003c/span\u003e\u003cspan address=\"10.1111/jep.14250\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAskell, A., Bai, Y., Chen, A., Drain, D., Ganguli, D., Henighan, T., Jones, A., Joseph, N., Mann, B., DasSarma, N.: A general language assistant as a laboratory for alignment. \u003cem\u003earXiv preprint arXiv:2112.00861\u003c/em\u003e. (2021). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2112.00861\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2112.00861\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAwad, E., Dsouza, S., Kim, R., Schulz, J., Henrich, J., Shariff, A., Bonnefon, J.-F., Rahwan, I.: The Moral Machine experiment. Nature. \u003cb\u003e563\u003c/b\u003e(7729), 59\u0026ndash;64 (2018). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41586-018-0637-6\u003c/span\u003e\u003cspan address=\"10.1038/s41586-018-0637-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBae, J., Kwon, S., Myeong, S.: Enhancing Software Code Vulnerability Detection Using GPT-4o and Claude-3.5 Sonnet: A Study on Prompt Engineering Techniques. \u003cem\u003eElectronics\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(13), 2657. (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.mdpi.com/2079-9292/13/13/2657\u003c/span\u003e\u003cspan address=\"https://www.mdpi.com/2079-9292/13/13/2657\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBansal, G., Chamola, V., Hussain, A., Guizani, M., Niyato, D.: Transforming Conversations with AI\u0026mdash;A Comprehensive Study of ChatGPT. Cogn. Comput. (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s12559-023-10236-2\u003c/span\u003e\u003cspan address=\"10.1007/s12559-023-10236-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBebeau, M.J., Thoma, S.J.: Intermediate Concepts and the Connection to Moral Education. Educational Psychol. Rev. \u003cb\u003e11\u003c/b\u003e(4), 343\u0026ndash;360 (1999). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1023/A:1022057316180\u003c/span\u003e\u003cspan address=\"10.1023/A:1022057316180\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBebeau, M.J., Thoma, S.J.: Guide for DIT-2. In: Minneapolis: Center for the Study of Ethical Development. University of Minnesota (2003)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBonagiri, V.K., Vennam, S., Govil, P., Kumaraguru, P., Gaur, M.: SaGE: Evaluating moral consistency in large language models. \u003cem\u003earXiv preprint arXiv:2402.13709\u003c/em\u003e. (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2402.13709\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2402.13709\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChu, Z., Ni, S., Wang, Z., Feng, X., Li, C., Hu, X., Xu, R., Yang, M., Zhang, W.: History, Development, and Principles of Large Language Models-An Introductory Survey. \u003cem\u003earXiv preprint arXiv:2402.06853\u003c/em\u003e. (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2402.06853\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2402.06853\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ede Winter, J., Dodou, D., Eisma, Y.B.: System 2 thinking in OpenAI's o1-preview model: Near-perfect performance on a mathematics exam. \u003cem\u003earXiv preprint arXiv:2410.07114\u003c/em\u003e. (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/computers13110278\u003c/span\u003e\u003cspan address=\"10.3390/computers13110278\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDevelopment, C.: f. t. S. o. E. \u003cem\u003eAbout the Intermediate Concepts Measures (ICM).\u003c/em\u003e. The University of Alabama. (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://ethicaldevelopment.ua.edu/about-the-intermediate-concepts-measures/\u003c/span\u003e\u003cspan address=\"https://ethicaldevelopment.ua.edu/about-the-intermediate-concepts-measures/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGan, W., Qi, Z., Wu, J., Lin, J.C.-W.: Large language models in education: Vision and opportunities. 2023 IEEE international conference on big data (BigData), (2023)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGanguli, D., Askell, A., Schiefer, N., Liao, T.I., Lukošiūtė, K., Chen, A., Goldie, A., Mirhoseini, A., Olsson, C., Hernandez, D.: The capacity for moral self-correction in large language models. \u003cem\u003earXiv preprint arXiv:2302.07459\u003c/em\u003e. (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2302.07459\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2302.07459\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGanguli, D., Hernandez, D., Lovitt, L., Askell, A., Bai, Y., Chen, A., Conerly, T., Dassarma, N., Drain, D., Elhage, N., Showk, S.E., Fort, S., Hatfield-Dodds, Z., Henighan, T., Johnston, S., Jones, A., Joseph, N., Kernian, J., Kravec, S., Clark, J.: \u003cem\u003ePredictability and Surprise in Large Generative Models\u003c/em\u003e Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, Seoul, Republic of Korea. (2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1145/3531146.3533229\u003c/span\u003e\u003cspan address=\"10.1145/3531146.3533229\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGehman, S., Gururangan, S., Sap, M., Choi, Y., Smith, N.A.: Realtoxicityprompts: Evaluating neural toxic degeneration in language models. \u003cem\u003earXiv preprint arXiv:2009.11462\u003c/em\u003e. (2020). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2009.11462\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2009.11462\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGungordu, N., Nabizadehchianeh, G., O\u0026rsquo;Connor, E., Ma, W., Walker, D.I.: Moral reasoning development: norms for Defining Issue Test-2 (DIT2). Ethics Behav. \u003cb\u003e34\u003c/b\u003e(4), 246\u0026ndash;263 (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/10508422.2023.2206573\u003c/span\u003e\u003cspan address=\"10.1080/10508422.2023.2206573\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHaenlein, M., Kaplan, A.: A Brief History of Artificial Intelligence: On the Past, Present, and Future of Artificial Intelligence. Calif. Manag. Rev. \u003cb\u003e61\u003c/b\u003e(4), 5\u0026ndash;14 (2019). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/0008125619864925\u003c/span\u003e\u003cspan address=\"10.1177/0008125619864925\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHager, P., Jungmann, F., Holland, R., Bhagat, K., Hubrecht, I., Knauer, M., Vielhauer, J., Makowski, M., Braren, R., Kaissis, G., Rueckert, D.: Evaluation and mitigation of the limitations of large language models in clinical decision-making. Nat. Med. \u003cb\u003e30\u003c/b\u003e(9), 2613\u0026ndash;2622 (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41591-024-03097-1\u003c/span\u003e\u003cspan address=\"10.1038/s41591-024-03097-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHan, H.: Potential benefits of employing large language models in research in moral education and development. J. Moral. Educ. 1\u0026ndash;16 (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/03057240.2023.2250570\u003c/span\u003e\u003cspan address=\"10.1080/03057240.2023.2250570\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIvanova, M.: AI, art and morality. AI Ethics. (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s43681-025-00735-3\u003c/span\u003e\u003cspan address=\"10.1007/s43681-025-00735-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJinnai, Y.: Does Cross-Cultural Alignment Change the Commonsense Morality of Language Models? \u003cem\u003earXiv preprint arXiv:2406.16316\u003c/em\u003e. (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2406.16316\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2406.16316\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJorgenson, C., Willems, J., Ozkes, A.I., Vanderelst, D.: Humans and robots are nearly ethically equivalent. AI Ethics. \u003cb\u003e5\u003c/b\u003e(3), 2919\u0026ndash;2928 (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s43681-024-00603-6\u003c/span\u003e\u003cspan address=\"10.1007/s43681-024-00603-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKerr, S.: \u003cem\u003eDeveloping and testing a teaching intermediate concept measure: A preliminary reliability and validity study\u003c/em\u003e University of Alabama Libraries]. (2018). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://ir.ua.edu/handle/123456789/5428\u003c/span\u003e\u003cspan address=\"http://ir.ua.edu/handle/123456789/5428\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKohlberg, L., Power, C.: Moral development, religious thinking, and the question of a seventh stage. Zygon: J. Relig. Sci. \u003cb\u003e16\u003c/b\u003e(3) (1981). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/j.1467-9744.1981.tb00417.x\u003c/span\u003e\u003cspan address=\"10.1111/j.1467-9744.1981.tb00417.x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMarcos, H.: Can large language models apply the law? \u003cem\u003eAI \u0026amp; SOCIETY\u003c/em\u003e. (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s00146-024-02105-9\u003c/span\u003e\u003cspan address=\"10.1007/s00146-024-02105-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcCarthy, J., Minsky, M.L., Rochester, N., Shannon, C.E.: A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence, August 31, 1955. AI Magazine. \u003cb\u003e27\u003c/b\u003e(4), 12 (2006). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1609/aimag.v27i4.1904\u003c/span\u003e\u003cspan address=\"10.1609/aimag.v27i4.1904\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNarvaez, D., Rest, J.: The four components of acting morally. Moral. Behav. moral. development: introduction. \u003cb\u003e1\u003c/b\u003e(1), 385\u0026ndash;400 (1995)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePeterson, M., G\u0026auml;rdenfors, P.: How to measure value alignment in AI. AI Ethics. \u003cb\u003e4\u003c/b\u003e(4), 1493\u0026ndash;1506 (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s43681-023-00357-7\u003c/span\u003e\u003cspan address=\"10.1007/s43681-023-00357-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePock, M., Ye, A., Moore, J.: LLMs grasp morality in concept. \u003cem\u003earXiv preprint arXiv:2311.02294\u003c/em\u003e. (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2311.02294\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2311.02294\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRamezani, A., Xu, Y.: Knowledge of cultural moral norms in large language models. \u003cem\u003earXiv preprint arXiv:2306.01857\u003c/em\u003e. (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2306.01857\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2306.01857\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRehman, U., Shah, M.U., Iqbal, F., Fatima, R.: Comparative analysis of moral decision-making and trust dynamics: human reasoning vs. ChatGPT-3 narratives. AI Ethics. \u003cb\u003e5\u003c/b\u003e(3), 2867\u0026ndash;2880 (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s43681-024-00605-4\u003c/span\u003e\u003cspan address=\"10.1007/s43681-024-00605-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRest, J., Narvaez, D., Bebeau, M.J., Thoma, S.J.: A Neo-Kohlbergian Approach: The DIT and Schema Theory. Educational Psychol. Rev. \u003cb\u003e11\u003c/b\u003e(4), 291\u0026ndash;324 (1999). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1023/A:1022053215271\u003c/span\u003e\u003cspan address=\"10.1023/A:1022053215271\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRest, J., Narvaez, D., Thoma, S.J., Bebeau, M.J.: DIT2: Devising and Testing A Revised Instrument of Moral Judgment. J. Educ. Psychol. \u003cb\u003e91\u003c/b\u003e, 644\u0026ndash;659 (1999). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1037/0022-0663.91.4.644\u003c/span\u003e\u003cspan address=\"10.1037/0022-0663.91.4.644\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRest, J., Thoma, S.J., Bebeau, M.J.: Postconventional moral thinking: A neo-Kohlbergian approach. Psychology (1999). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.4324/9781410603913\u003c/span\u003e\u003cspan address=\"10.4324/9781410603913\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRest, J.R.: Background: Theory and research. In: Moral development in the professions, pp. 13\u0026ndash;38. Psychology (1994a)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRest, J.R.: Moral development in the professions: Psychology and applied ethics. Psychology (1994b)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRest, J.R., Narvaez, D., Thoma, S.J., Bebeau, M.J.: A Neo-Kohlbergian Approach to Morality Research. J. Moral. Educ. \u003cb\u003e29\u003c/b\u003e(4), 381\u0026ndash;395 (2000). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/713679390\u003c/span\u003e\u003cspan address=\"10.1080/713679390\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRichardson, H.S.: Moral reasoning. (2003)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoberts, F.C., Park, O.: Intelligent Computer-Assisted Instruction: An Explanation and Overview. Educational Technol. \u003cb\u003e23\u003c/b\u003e(12), 7\u0026ndash;12 (1983). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://www.jstor.org/stable/44426785\u003c/span\u003e\u003cspan address=\"http://www.jstor.org/stable/44426785\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRodr\u0026iacute;guez-L\u0026oacute;pez, B., Rueda, J.: Artificial moral experts: asking for ethical advice to artificial intelligent assistants. AI Ethics. \u003cb\u003e3\u003c/b\u003e(4), 1371\u0026ndash;1379 (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s43681-022-00246-5\u003c/span\u003e\u003cspan address=\"10.1007/s43681-022-00246-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoumeliotis, K.I., Tselikas, N.D.: ChatGPT and Open-AI Models: A Preliminary Review. Future Internet. \u003cb\u003e15\u003c/b\u003e(6), 192 (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.mdpi.com/1999-5903/15/6/192\u003c/span\u003e\u003cspan address=\"https://www.mdpi.com/1999-5903/15/6/192\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSebo, J., Long, R.: Moral consideration for AI systems by 2030. AI Ethics. \u003cb\u003e5\u003c/b\u003e(1), 591\u0026ndash;606 (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s43681-023-00379-1\u003c/span\u003e\u003cspan address=\"10.1007/s43681-023-00379-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSengupta, S., Flathmann, C., Schelble, B., Lyons, J.B., McNeese, N.: An analysis of ethical rationales and their impact on the perceived moral persona of AI teammates. AI Ethics. \u003cb\u003e5\u003c/b\u003e(3), 2667\u0026ndash;2679 (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s43681-024-00515-5\u003c/span\u003e\u003cspan address=\"10.1007/s43681-024-00515-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSimmons, G.: Moral mimicry: Large language models produce moral rationalizations tailored to political identity. \u003cem\u003earXiv preprint arXiv:2209.12106\u003c/em\u003e. (2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2209.12106\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2209.12106\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSlavkovik, M.: Automating moral reasoning. \u003cem\u003eInternational Research School in Artificial Intelligence in Bergen (AIB 2022)\u003c/em\u003e. (2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.4230/OASIcs.AIB.2022.6\u003c/span\u003e\u003cspan address=\"10.4230/OASIcs.AIB.2022.6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStenseke, J.: Artificial virtuous agents: from theory to machine implementation. AI Soc. \u003cb\u003e38\u003c/b\u003e(4), 1301\u0026ndash;1320 (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s00146-021-01325-7\u003c/span\u003e\u003cspan address=\"10.1007/s00146-021-01325-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStenseke, J.: Artificial virtuous agents in a multi-agent tragedy of the commons. AI Soc. \u003cb\u003e39\u003c/b\u003e(3), 855\u0026ndash;872 (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s00146-022-01569-x\u003c/span\u003e\u003cspan address=\"10.1007/s00146-022-01569-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSullivan, Y.W., Fosso Wamba, S.: Moral Judgments in the Age of Artificial Intelligence. J. Bus. Ethics. \u003cb\u003e178\u003c/b\u003e(4), 917\u0026ndash;943 (2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10551-022-05053-w\u003c/span\u003e\u003cspan address=\"10.1007/s10551-022-05053-w\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSwoboda, T., Lauwaert, L.: Can artificial intelligence embody moral values? AI Ethics. (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s43681-025-00662-3\u003c/span\u003e\u003cspan address=\"10.1007/s43681-025-00662-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSzutta, A.: Artificial intelligence as a moral mentor. J. Moral. Educ. 1\u0026ndash;19 (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/https://doi.org/10.1080/03057240.2025.2475539\u003c/span\u003e\u003cspan address=\"10.1080/03057240.2025.2475539\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTakemoto, K.: The moral machine experiment on large language models. Royal Soc. open. Sci. \u003cb\u003e11\u003c/b\u003e(2), 231393 (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2411.06790\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2411.06790\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTanmay, K., Khandelwal, A., Agarwal, U., Choudhury, M.: Probing the Moral Development of Large Language Models through Defining Issues Test. \u003cem\u003earXiv e-prints\u003c/em\u003e, arXiv: \u003cdiv class=\"ExternalRefDOI\"\u003e2309.13356\u003c/div\u003e. (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/https://ui.adsabs.harvard.edu/link_gateway/2023arXiv230913356T/doi:10.48550/arXiv.2309.13356\u003c/span\u003e\u003cspan address=\"https://doi.org/https://ui.adsabs.harvard.edu/link_gateway/2023arXiv230913356T/doi:10.48550/arXiv.2309.13356\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTeam, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.-B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A.: Gemini: a family of highly capable multimodal models. \u003cem\u003earXiv preprint arXiv:2312.11805\u003c/em\u003e. (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2312.11805\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2312.11805\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThoma, S.: Measuring Moral Judgments from a Neo-Kohlbergian Perspective. \u003cem\u003ePaper presented at Can Virtue Be Measured?\u003c/em\u003e (2014)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThoma, S., Derryberry, W.P., Crowson, H.M.: Describing and testing an intermediate concept measure of adolescent moral thinking. Eur. J. Dev. Psychol. \u003cb\u003e10\u003c/b\u003e(2), 239\u0026ndash;252 (2013). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/17405629.2013.766129\u003c/span\u003e\u003cspan address=\"10.1080/17405629.2013.766129\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThoma, S.J., Bebeau, M., Bolland, A.: The role of moral judgment in context-specific professional decision making. In \u003cem\u003eGetting Involved\u003c/em\u003e (pp. 147\u0026ndash;160). Brill. (2008). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1163/9789087906368_011\u003c/span\u003e\u003cspan address=\"10.1163/9789087906368_011\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThoma, S.J., Dong, Y.: The Defining Issues Test of moral judgment development. Behav. Dev. Bull. \u003cb\u003e19\u003c/b\u003e(3), 55\u0026ndash;61 (2014). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1037/h0100590\u003c/span\u003e\u003cspan address=\"10.1037/h0100590\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThoma, S.J., Walker, D.I., Chen, Y.H., Frichand, A., Moulin-Stożek, D., Kristj\u0026aacute;nsson, K.: Adolescents' application of the virtues across five cultural contexts. Dev. Psychol. \u003cb\u003e55\u003c/b\u003e(10), 2181\u0026ndash;2192 (2019). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1037/dev0000770\u003c/span\u003e\u003cspan address=\"10.1037/dev0000770\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVolkman, R., Gabriels, K.: AI Moral Enhancement: Upgrading the Socio-Technical System of Moral Engagement. Sci Eng. Ethics. \u003cb\u003e29\u003c/b\u003e(2), 11 (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s11948-023-00428-2\u003c/span\u003e\u003cspan address=\"10.1007/s11948-023-00428-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWalker, D.I., Thoma, S.J., Arthur, J.: Assessing ethical reasoning among junior British Army officers using the Army Intermediate Concept Measure (AICM). J. Military Ethics. \u003cb\u003e20\u003c/b\u003e(1), 2\u0026ndash;20 (2021). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/15027570.2021.1895965\u003c/span\u003e\u003cspan address=\"10.1080/15027570.2021.1895965\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWangsa, K., Karim, S., Gide, E., Elkhodr, M.: A Systematic Review and Comprehensive Analysis of Pioneering AI Chatbot Models from Education to Healthcare: ChatGPT, Bard, Llama, Ernie and Grok. \u003cem\u003eFuture Internet\u003c/em\u003e, \u003cem\u003e16\u003c/em\u003e(7), 219. (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.mdpi.com/1999-5903/16/7/219\u003c/span\u003e\u003cspan address=\"https://www.mdpi.com/1999-5903/16/7/219\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu, S., Peng, Z., Du, X., Zheng, T., Liu, M., Wu, J., Ma, J., Li, Y., Yang, J., Zhou, W.: A Comparative Study on Reasoning Patterns of OpenAI's o1 Model. \u003cem\u003earXiv preprint arXiv:2410.13639\u003c/em\u003e. (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2410.13639\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2410.13639\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXing, W., Nixon, N., Crossley, S., Denny, P., Lan, A., Stamper, J., Yu, Z.: The Use of Large Language Models in Education. Int. J. Artif. Intell. Educ. (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s40593-025-00457-x\u003c/span\u003e\u003cspan address=\"10.1007/s40593-025-00457-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXu, Z., Sengar, N., Chen, T., Chung, H., Oviedo-Trespalacios, O.: Where is morality on wheels? Decoding large language model (LLM)-driven decision in the ethical dilemmas of autonomous vehicles. Travel Behav. Soc. \u003cb\u003e40\u003c/b\u003e, 101039 (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/https://doi.org/10.1016/j.tbs.2025.101039\u003c/span\u003e\u003cspan address=\"10.1016/j.tbs.2025.101039\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYan, L., Sha, L., Zhao, L., Li, Y., Martinez-Maldonado, R., Chen, G., Li, X., Jin, Y., Gašević, D.: Practical and ethical challenges of large language models in education: A systematic scoping review. Br. J. Edu. Technol. \u003cb\u003e55\u003c/b\u003e(1), 90\u0026ndash;112 (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/https://doi.org/10.1111/bjet.13370\u003c/span\u003e\u003cspan address=\"10.1111/bjet.13370\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZawacki-Richter, O., Mar\u0026iacute;n, V.I., Bond, M., Gouverneur, F.: Systematic review of research on artificial intelligence applications in higher education\u0026ndash;where are the educators? Int. J. Educational Technol. High. Educ. \u003cb\u003e16\u003c/b\u003e(1), 1\u0026ndash;27 (2019). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/https://doi.org/10.1186/s41239-019-0171-0\u003c/span\u003e\u003cspan address=\"10.1186/s41239-019-0171-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, N., Sun, Z., Xie, Y., Wu, H., Li, C.: The latest version ChatGPT powered by GPT-4o: what will it bring to the medical field? Int. J. Surg. (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1097/js9.0000000000001754\u003c/span\u003e\u003cspan address=\"10.1097/js9.0000000000001754\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao, W., Zhou, K., Junyi, L., Tianyi, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., Dong, Z., Du, Y., Yang, C., Chen, Y., Jiang, J., Ren, R., Li, Y., Tang, X., Liu, Z., Wen, J.-R.: \u003cem\u003eA Survey of Large Language Models\u003c/em\u003e. (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2303.18223\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2303.18223\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou, J., Hu, M., Li, J., Zhang, X., Wu, X., King, I., Meng, H.: Rethinking Machine Ethics\u0026ndash;Can LLMs Perform Moral Reasoning through the Lens of Moral Theories? \u003cem\u003earXiv preprint arXiv:2308.15399\u003c/em\u003e. (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2308.15399\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2308.15399\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhui, L., Fenghe, L., Xuehu, W., Qining, F., Wei, R.: Ethical Considerations and Fundamental Principles of Large Language Models in Medical Education: Viewpoint. J. Med. Internet Res. \u003cb\u003e26\u003c/b\u003e, e60083 (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.2196/60083\u003c/span\u003e\u003cspan address=\"10.2196/60083\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZubiaga, A.: Natural language processing in the era of large language models [Specialty Grand Challenge]. \u003cem\u003eFrontiers in Artificial Intelligence\u003c/em\u003e, \u003cem\u003e6\u003c/em\u003e. (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/frai.2023.1350306\u003c/span\u003e\u003cspan address=\"10.3389/frai.2023.1350306\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Large Language Models (LLMs), Moral Reasoning, Defining Issues Test (DIT-2), Intermediate Concept Measure (ICM)","lastPublishedDoi":"10.21203/rs.3.rs-6823916/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6823916/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThis study investigates how various large language models (LLMs) generate responses to moral reasoning dilemmas. It specifically examines LLM-generated responses using the Defining Issues Test (DIT-2) and the Intermediate Concepts Measure (ICM) for Educational Leaders. Using a neo-Kohlbergian approach to moral reasoning, the study evaluates responses from multiple LLM platforms: ChatGPT-3.5, ChatGPT-4, ChatGPT-4O, Grok Premium Plus, Claude 3.5 Sonnet, Gemini, and Gemini Advanced. For DIT-2, Claude learns to prioritize the highest post-conventional moral reasoning score and N2 score (P-score 72, N2 score 71.10), followed by Gemini Advanced (P-score 64, N2 score 60.31) and Gemini (P-score 58, N2 score 52.11). Other LLMs performed as follows: Grok (P-score 48, N2 score 47.98), ChatGPT-4O (P-score 44, N2 score 55.07), ChatGPT-4 (P-score 44, N2 score 46.53), and ChatGPT-3.5 (P-score 18, N2 score 36.20). For the ICM Educational Leaders version, Gemini Advanced had the highest total ICM score of 0.90, followed by Claude 3.5 Sonnet and Gemini (both 0.86), ChatGPT-4O and ChatGPT-4 (both 0.78), Grok (0.61), and ChatGPT-3.5 (0.32). The findings indicate that some LLMs can generate responses consistent with sophisticated moral reasoning patterns, producing scores comparable to or exceeding graduate-level human participants (whose P-scores typically range from 38.5 to 42.3) and provide a methodological framework consisting of standardized assessment protocols and comparative analysis techniques for larger-scale research to improve our understanding of AI's potential in moral reasoning.\u003c/p\u003e","manuscriptTitle":"Exploring Large Language Models' Responses to Moral Reasoning Dilemmas","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-17 15:45:23","doi":"10.21203/rs.3.rs-6823916/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"c2ca716e-d9fa-4768-a8e5-ec414cdca6dd","owner":[],"postedDate":"June 17th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-08-16T16:08:36+00:00","versionOfRecord":[],"versionCreatedAt":"2025-06-17 15:45:23","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6823916","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6823916","identity":"rs-6823916","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}