AI's Rational Empathy Promotes Reconciliation in Conflict: Evidence from Behavioral Experiments, Linguistic Analysis, and Topic Modeling

AI's Rational Empathy Promotes Reconciliation in Conflict: Evidence from Behavioral Experiments, Linguistic Analysis, and Topic Modeling | 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 AI's Rational Empathy Promotes Reconciliation in Conflict: Evidence from Behavioral Experiments, Linguistic Analysis, and Topic Modeling Tongtong Ye, Haoliang Mai, Hongyu Wang, Haodong Su This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7699927/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 With the rise of large language models (LLMs), LLM-driven artificial intelligence role-playing (AI-RP) has emerged as a promising tool for resolving interpersonal conflict. However, the differential effectiveness of its "Rational Empathy" compared to human empathy remains unclear. This study investigates this issue through two experiments. Study 1 examined the effects of different AI response styles, while Study 2 directly compared interventions by an AI and a human counsellor, incorporating linguistic analysis (LIWC and BERTopic) to investigate the underlying mechanisms. Results revealed that while AI-RP effectively improved conflict resolution outcomes, this effect was context-dependent and did not generalize. Crucially, on the key metric of communication intention, the AI was significantly superior to the human counsellor. Linguistic analysis indicated that the AI’s responses were more focused on functional, problem-solving approaches, whereas the counsellor’s focused more on affective and relational aspects. This research demonstrates that an AI can act as a "cognitive scaffold" in conflicts. Its unique advantage stems from an efficient, problem-oriented "Rational Empathy" that signals the viability of communication, offering a new perspective for future human-AI collaborative interventions. Psychology large language models AI Human Counsellor Empathy Conflict Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Introduction While the use of chatbots for obtaining basic, factual information has long been commonplace(Adamopoulou & Moussiades, 2020), the advent of Large Language Models (LLMs) since 2022, spearheaded by OpenAI's ChatGPT, has profoundly reshaped the paradigm of human-chatbot interaction(Kalla et al., 2023). Diverging from the utilitarian expectations placed on traditional virtual assistants, user expectations for LLM-based chatbots are increasingly shifting towards the fulfillment of emotional and psychological needs(Achiam et al., 2023). This shift highlights the immense potential of LLMs as tools for fostering human psychological well-being and social competencies. Among the myriad of human social experiences, interpersonal conflict represents a particularly pervasive and challenging domain. The effective management of such conflicts is crucial for maintaining mental health, fostering interpersonal harmony, and ensuring social stability(Thoits, 2011). LLMs are now being applied as mediators to address real-world interpersonal conflicts in areas such as automated negotiation, group decision-making, argumentation, preference aggregation, and human-computer interaction(Aydoğan et al., 2021). This application persists despite the well-documented phenomenon of "algorithm aversion"(Feldkamp et al., 2023; Kim & Peng, 2024), which encapsulates public skepticism about whether an impersonal AI can genuinely understand and navigate complex human emotions. Such skepticism is particularly pronounced in domains heavily reliant on empathy. Nevertheless, the use of LLMs as auxiliary tools for emotional support and conflict mediation is becoming increasingly common(Hsu & Chaudhary, 2023). At the core of effective interpersonal conflict mediation lies empathy—the ability to experience the world from another's perspective(Chaturvedi, 2024). In conflict situations, empathy can reduce aggressive attitudes(Rosler et al., 2015) and lead to lower levels of escalation(Klimecki, 2019). Through extensive training on vast datasets, LLMs enable computers to express empathy by analyzing and responding to user inputs(Adam et al., 2021). However, translating an LLM's capacity for empathy into an effective conflict intervention requires a suitable medium of implementation. Drawing on symbolic interactionism, Mead(Mead, 2015)proposed that the self develops through the continuous interaction between the "I" and the "Me," and that we come to understand others' perspectives and attitudes through role-taking. Reinforcing this, extensive research in social psychology has confirmed that role-playing is a powerful intervention technique. By immersing individuals in simulated conflict scenarios, it effectively promotes perspective-taking, elicits cognitive and emotional responses, and ultimately improves their conflict resolution strategies(Lutgen-Sandvik & Sypher, 2009; Taguchi, 2023). Role-playing thus offers an ideal research paradigm for the present study: it serves as both an established technique for modifying conflict responses and a controlled, safe environment for precisely comparing the mediating effects of an LLM acting as a third party. Previous research has established that individuals typically seek external emotional and social support when facing interpersonal conflicts or stressful events(Thoits, 2011). Positive and empathic responses not only buffer negative emotions but also facilitate constructive communication and conflict resolution(Pagani et al., 2019). Lutgen-Sandvik and Sypher(Lutgen-Sandvik & Sypher, 2009) categorized conflict coping styles into two types: destructive approaches, such as blaming, attacking, and retaliating, and constructive approaches, which include understanding, perspective-taking, and constructive communication. The present study is the first to systematically investigate the effectiveness of AI-enabled role-playing (AI-RP) in influencing participants' interpersonal conflict resolution strategies through a behavioral experiment. Accordingly, we propose our first hypothesis (H1): Participants who engage in role-playing with an AI will adopt more constructive conflict resolution strategies than those in a no-interaction control group. Specifically, participants interacting with an AI exhibiting a positive, empathic response style will demonstrate significantly more constructive approaches (i.e., higher intentions for understanding and communication, and lower intention for retaliation) compared to both participants interacting with a negative-response AI and those in the control group. A deeper question then arises: when both an AI and a professionally trained human counselor provide empathy during role-playing, which approach is more effective in improving conflict resolution outcomes? Empathy is a multidimensional construct, comprising both affective and cognitive dimensions(Davis, 1983). The empathy of a human counselor is rooted in genuine emotional experience, a significant strength that can also be a limitation. For instance, phenomena such as transference and countertransference may lead to biased judgments, while sustained emotional involvement can result in compassion fatigue(Figley, 2002; Levy & Scala, 2012). In contrast, the empathy exhibited by an AI is what we term "Rational Empathy"—a form of empathy derived from data-driven pattern recognition and language generation. Devoid of genuine emotional experience, it is immune to emotional fluctuations, biases, or fatigue, allowing it to provide more stable, consistent, and controllable empathic responses(Adam et al., 2021). This sets up a key theoretical tension: the embodied, authentic emotional empathy of a human counselor versus the disembodied, computational "Rational Empathy" of an AI. On one hand, a genuine emotional connection may be the critical catalyst for user change. On the other hand, stable and unbiased rational support might offer a safer conversational space in highly emotional conflicts. Existing theories do not offer a clear prediction as to which approach is superior. Therefore, the second core exploratory research question of this study is as follows: In the context of conflict role-playing, which interaction—with an AI or with a human counselor—is more effective at promoting an individual's constructive behaviors in subsequent conflict situations? We will address this question by experimentally comparing the effectiveness of these two intervention modalities. Furthermore, we will conduct an in-depth analysis of the dialogue texts to investigate the potential mechanistic differences between the AI's "Rational Empathy" and the human counselor's "Emotional Empathy." To systematically address the aforementioned research questions and test our hypotheses, we designed a series of two studies. Study 1, a behavioral experiment, was designed to investigate the impact of different AI response styles (positive, negative, and a no-interaction control) during role-playing on participants' subsequent conflict behaviors. Study 2 directly compared positive role-playing interactions with an AI versus a human counselor to explore the differential effects of "Rational Empathy" and "Emotional Empathy" on conflict resolution. Furthermore, to delve into the underlying psychological mechanisms, we employed a multi-method approach, combining behavioral data with LIWC-based linguistic analysis and BERTopic-based topic modeling. This approach allows us to reveal the multifaceted impacts of different interaction modalities on individuals' cognition, emotions, and behaviors. Study 1 Method Participants An a priori power analysis was conducted using G*Power (Version 3.1) to determine the required sample size. The analysis was set for a MANOVA: Global effects, with parameters of α = 0.05, power (1-β) = 0.80, and a medium effect size (f²(V) = 0.08). The results indicated a minimum required sample of 90 participants. We initially recruited 99 university students. Twenty participants were excluded as they failed to complete the entire experimental procedure, resulting in a final sample of 79 individuals. Participants were randomly assigned to one of three groups: a positive-response group (PR; n = 24, 8 males, 16 females; M = 19.08, SD = 0.78), a negative-response group (NR; n = 31, 18 males, 13 females; M = 19.26, SD = 1.06), or a no-interaction control group (CON; n = 24, 9 males, 15 females; M = 20.63, SD = 1.93). During the experiment, participants were blind to their assigned condition and the study's true hypotheses. This research was approved by the Institutional Review Board of Anhui University of Science and Technology. All participants provided written informed consent prior to the experiment and were informed of their right to withdraw at any time. Design A single-factor, between-subjects design was employed. The independent variable was the AI's response style during role-playing, with three levels: positive, negative, and a no-interaction control. The dependent variables were participants' intentions of retaliation, understanding, and communication. This study aimed to examine the effect of the AI's response style on participants' reactions in conflict scenarios and to test the generalization of this effect. To assess generalization, participants' responses were measured in two contexts: (1) a repeated context, identical to the one used in the role-play, and (2) a parallel context, which was thematically similar but distinct from the role-play scenario. Materials and Measures Conflict Scenarios . The scenario for the role-play (and repeated context measure) described a conflict regarding reciprocal responsibilities within a three-person study group. The protagonist had fulfilled their duty of saving a library seat with a power outlet for the group on three occasions, whereas a partner had failed to do so on two of their designated days without prior notice. This negligence resulted in the protagonist being unable to secure an adequate seat for a group discussion, ultimately disrupting their preparation when their device's battery depleted. The parallel context described a conflict in a course group project: the protagonist, as the group leader, publicly pointed out that a member (A) had repeatedly failed to complete their assigned tasks. Subsequently, Member A quit the group chat, locked the shared files, and submitted their own work independently, thereby hindering the group's progress. Dependent Measures . Participants' behavioral intentions were measured using four items adapted from prior research on conflict resolution(Jia, 2014). All items were rated on a 5-point Likert scale (1 = Strongly Disagree, 5 = Strongly Agree). Two items assessed retaliation intention (e.g., "Find a similar opportunity to make him/her feel the same way," "Spread what he/she did to other classmates"). One item assessed understanding (e.g., "He/she was just acting according to his/her personality; I can understand him/her"). One item assessed communication intention (e.g., "Find a suitable opportunity to have an effective conversation with him/her"). Communication and AI Tools . The interaction was conducted via the instant messaging software QQ. To standardize the interaction and control the AI's persona, we employed a "Wizard of Oz" technique. A research assistant, unknown to the participant, acted as an intermediary, relaying messages between the participant and the AI model. This setup also allowed us to control the timing and pace of the AI's responses. Following a pilot evaluation of several Large Language Models (LLMs), "Doubao" was selected as the AI tool for its superior performance in adhering to the complex instructional prompts required for the positive and negative response conditions. Procedure The experiment was conducted in a standardized laboratory setting over three consecutive days, with conditions counterbalanced across time slots. After providing informed consent, participants were randomly assigned to one of the three conditions (see Figure 1 for a detailed flowchart). Participants were instructed to begin a conversation via QQ with a "partner" (the AI) about the library seat-saving conflict scenario. They were asked to send the first message, after which the interaction proceeded for five turns per participant (totaling 10 messages). In the positive-response (PR) condition, the AI was prompted to provide empathic and understanding responses. In the negative-response (NR) condition, the AI was prompted to be negative and unsupportive. Participants in the control (CON) group did not interact with the AI; instead, after reading the scenario, they were instructed to imagine being in the situation for three minutes. Immediately following the interaction (or imagination task), participants completed a manipulation check, rating their engagement in the role-play on a 5-point scale. Subsequently, they completed an online questionnaire containing the dependent measures for both the repeated and parallel contexts. Finally, all participants were fully debriefed regarding the true purpose of the study and the use of AI. Data Analysis A one-way multivariate analysis of variance (MANOVA) was conducted using JASP (Version 0.95) to examine the overall effect of the AI's response style on the three dependent variables. Pillai's Trace was selected as the test statistic due to its robustness. If the MANOVA yielded a significant result, a series of one-way ANOVAs were performed on each dependent variable to identify specific effects. Significant main effects were followed up with Tukey's HSD post-hoc tests for pairwise comparisons. The significance level was set at p < 0.05. All data are reported as M ± SE . Result Given that the dependent variables of interest—retaliation, understanding, and communication—were theoretically related, a one-way multivariate analysis of variance (MANOVA) was conducted to test for an overall effect of the AI's response style on the linear combination of these variables. The results of the MANOVA are presented in Table 1. The analysis revealed a significant multivariate main effect of condition in the repeated scenario, indicating a significant difference among the groups on the combined dependent variables. However, this multivariate effect was not significant in the parallel scenario. To follow up on the significant multivariate effect in the repeated scenario, a series of one-way ANOVAs were conducted separately for each of the three dependent variables. These analyses revealed a significant main effect of condition on retaliation, F (2, 76) = 3.76, p = 0.028, ω² = 0.07; understanding, F (2, 76) = 7.17, p = 0.001, ω² = 0.14; and communication, F (2, 76) = 7.31, p = 0.001, ω² = 0.14. Pairwise comparisons using Tukey's HSD post-hoc tests clarified the nature of these effects (see Figures 2A-C). For the retaliation variable, the positive-response (PR) group reported significantly lower intentions than the control (CON) group, while no other pairwise comparisons were significant (Fig.2A). A different pattern emerged for understanding, where both the PR and negative-response (NR) groups demonstrated significantly higher intentions than the CON group, with no significant difference found between the two AI conditions (Fig.2B). Finally, regarding the intention to communicate, the PR group scored significantly higher than both the NR and CON groups, between which there was no significant difference (Fig. 2C). The distributions of individual data points for each group are presented in Figures 2D-F. Discussion1 The results of Study 1 demonstrate that engaging in role-playing with an AI in a simulated conflict scenario significantly enhanced participants' intentions to understand and communicate, while reducing their retaliatory tendencies. This finding supports our first hypothesis (H1). However, this positive effect did not generalize to the novel parallel context. This suggests that AI-enabled role-playing (AI-RP) may function as a context-dependent intervention, whose effects diminish once the immediate intervention context is removed. This phenomenon can be interpreted through the lens of Vygotsky's theory of cognitive scaffolding(Vygotsky & Cole, 1978). Interpersonal conflicts are inherently situations of high cognitive load(Sweller, 1988). Within the structured and simulated nature of the role-playing task, participants' cognitive load may be temporarily alleviated, enabling them to allocate more cognitive resources towards understanding their partner's perspective. In this context, the AI acts as a form of external support, providing a "cognitive scaffold" that temporarily reduces the burden of managing the high-load conflict situation. However, once this scaffold is removed and participants are confronted with a new parallel context, their cognitive load likely increases, causing them to revert to their habitual, more automatic modes of conflict resolution. A deeper analysis of the different AI response styles reveals a nuanced pattern. While the positive-response (PR) group exhibited significantly higher communication intentions than the negative-response (NR) group, the two groups did not differ significantly on the dimensions of understanding and retaliation. Notably, the NR group still demonstrated significantly better outcomes on these two dimensions compared to the no-interaction control group. These results point toward two distinct mechanisms underlying the AI-RP intervention. The first mechanism is the activation of perspective-taking. Regardless of the AI's response style, the role-playing exercise itself provides a continuous, external interaction partner, which appears to compel participants to engage in perspective-taking(Galinsky & Moskowitz, 2000). The necessity of responding to the AI partner forces individuals to move beyond their own preconceived biases and consider the other's point of view, a process known to reduce animosity and enhance understanding(Galinsky et al., 2005). Thus, a fundamental function of AI-RP is to provide a safe interactive space that prompts cognitive reappraisal, thereby reducing retaliatory intentions and increasing understanding. The second mechanism is the provision of emotional support, where the AI's positive empathy plays an additional, distinct role. Initiating communication is often perceived as having high interpersonal risk. By experiencing understanding and acceptance from the positive AI, participants' sense of psychological safety is enhanced, which can significantly lower the perceived risks associated with communication(Fredrickson, 2001). This, in turn, motivates them to engage in more constructive communication behaviors. In summary, Study 1 provides a preliminary validation of the feasibility of AI-enabled role-playing (AI-RP) for improving conflict resolution strategies, highlighting its distinct roles in activating perspective-taking and providing emotional support. However, these findings raise a deeper theoretical question: Do the "algorithmic empathy" of an AI and the "embodied empathy" of a human differ fundamentally in nature, and does this potential difference impact the intervention's effectiveness? While some scholars argue that an AI cannot achieve genuine human empathy due to its lack of authentic emotional experience(Harel, 2004), its "Rational Empathy" may, conversely, offer unique advantages owing to its impartiality and consistency(Adam et al., 2021). Therefore, Study 2 was designed to directly compare role-playing with an AI versus a human counselor, aiming to investigate which form of empathy is more conducive to promoting constructive conflict resolution. Furthermore, through the analysis of dialogue texts, we will delve into the distinct psychological mechanisms underlying these two intervention modalities. This will provide empirical evidence for understanding the differential utility of "Rational Empathy" versus "Emotional Empathy" in practical application. Study 2 Study 2a Method Participants An a priori power analysis was conducted using G*Power (Version 3.1) to determine the required sample size. The analysis was set for a MANOVA: Global effects, with parameters of α = 0.05, power (1-β) = 0.80, and a medium effect size (f²(V) = 0.08), which indicated a minimum required sample of 90 participants. We initially recruited 100 university students. Seventeen participants were excluded for failing to complete the full experimental procedure, resulting in a final sample of 83 individuals. Participants were randomly assigned to one of three groups: a human counselor group (n = 29, 8 males, 16 females; M = 20.24, SD = 1.66), an AI group (n = 26, 7 males, 19 females; M = 19.85, SD = 1.22), or a no-interaction control group (CON; n = 28, 7 males, 21 females; M = 20.46, SD = 1.32). All experimental and ethical procedures were consistent with those of Study 1. Design A single-factor, between-subjects design was employed. The independent variable was the type of role-playing partner, with three levels: a human counsellor(Counsellor), an artificial intelligence (AI), and a no-interaction control(Control). The dependent variables were participants' intentions of retaliation, understanding, and communication. This study aimed to directly compare the effects of role-playing with a human counselor versus an AI on participants' conflict resolution intentions under a positive-response condition. Materials and Measures Conflict Scenario . The conflict scenario used in this study involved an online multiplayer game: the protagonist publicly pointed out a teammate's (F's) tactical error that led to the team's loss. In response, Teammate F publicly posted the protagonist's previous term paper plagiarism report in a 200-person university major group chat, questioning their academic integrity. Dependent Measures . The measures for retaliation, understanding, and communication intentions were identical to those used in Study 1, employing a 5-point Likert scale. Communication and AI Tools . As in Study 1, interactions were conducted via QQ using the "Wizard of Oz" technique. The "Doubao" model was again used as the AI tool. Human Counsellor. Interactions in the counselor group were conducted by a senior licensed psychotherapist with over 10 years of professional training and 7 years of clinical experience, accumulating nearly 3,000 hours of client sessions and approximately 500 hours of supervised practice. Procedure The overall experimental procedure was similar to that of Study 1, with three key differences:(1) To gather richer dialogue data for subsequent analysis (Study 2b), the role-playing interaction was extended to six turns per participant (12 messages in total); (2) Both the AI and the human counselor were instructed to provide positive, empathic responses, following identical intervention guidelines to ensure stylistic consistency; (3) Unlike Study 1, this study did not include a parallel context, aiming to focus more directly on the immediate effects of the two intervention modalities in a specific scenario. Consequently, the dependent variables were measured only in relation to the role-playing context itself. The procedure for the control (CON) group was identical to that in Study 1. All participants were fully debriefed at the conclusion of the experiment. Data Analysis The quantitative data analysis approach for Study 2a (MANOVA followed by ANOVAs) was identical to that of Study 1. In addition, the dialogue texts from the two experimental groups (AI and counsellor) were coded for conflict resolution outcomes, categorized into three types: (1) Conflict Resolved (e.g., participant expressed positive emotions and willingness to reconcile); (2) Conflict Partially Resolved (e.g., participant still expressed some negativity but was hopeful about the future relationship); and (3) Conflict Unresolved (e.g., participant explicitly refused reconciliation). To ensure objectivity, two independent raters, blind to the experimental conditions, coded the outcomes. Inter-rater reliability was assessed using Cohen's Kappa, which indicated a substantial level of agreement (κ = 0.727, p < 0.001, 95% CI [0.577, 0.878]). Result The results of the one-way MANOVA are presented in Table 2. The analysis revealed a significant multivariate main effect of the role-playing partner on the combined dependent variables of retaliation, understanding, and communication. Subsequent one-way ANOVAs were conducted for each dependent variable. These analyses revealed a significant main effect of condition on retaliation, F (2, 80) = 7.27, p = 0.001, ω² = 0.13; understanding, F (2, 80) = 8.25, p < 0.001, ω² = 0.15; and communication, F (2, 80) = 8.24, p < 0.001, ω² = 0.15. Pairwise comparisons using Tukey's HSD post-hoc tests clarified these effects (see Figures 4A-C). For retaliation, both the AI and Counsellor groups reported significantly lower intentions than the Control group, with no significant difference between the AI and Counsellor groups themselves (Fig.4A). A similar pattern was observed for understanding, where the AI and Counsellor groups both reported significantly higher intentions than the Control group, and again did not differ from each other (Fig.4B). For communication, however, the AI group scored significantly higher than both the Counsellor and Control groups, between which there was no significant difference (Fig.4C). In addition to the scaled measures, we analyzed the categorical conflict resolution outcomes based on the coded dialogue texts (see Figure 4D). The results were broadly consistent with the post-hoc findings. A slightly higher percentage of participants in the AI group (64.52%) reached a "Conflict Resolved" outcome compared to the Counsellor group (60.71%). The proportions of participants in the "Partially Resolved" and "Unresolved" categories were comparable across the two groups. Study 2b Method Participants and Corpus The present analysis was conducted on the dialogue text data generated by the 55 participants in the AI (n = 26) and Counsellor (n = 29) groups from Study 2a. The complete dialogue corpus consisted of 660 conversational turns (55 participants × 6 turns × 2 interlocutors). For comparative analysis, the corpus was divided into two sub-corpora: participant-generated text and partner-generated text (from either the AI or the human counsellor). Data Analysis Strategy To investigate differences in the linguistic style and content between the AI and human counsellor interactions, we employed two computational methods: linguistic feature analysis and topic modeling. All analyses were performed in a Python 3.10 environment using Jupyter Notebooks. Text Preprocessing ：The text preprocessing pipeline was adapted for the specific requirements of each analytical method. For the linguistic feature analysis (LIWC), the text underwent several steps: removal of non-textual content (e.g., emojis), punctuation removal, Chinese word segmentation using the jieba library, and stop-word removal. For the topic modeling (BERTopic), only non-textual content was removed to preserve the complete semantic structure of the sentences. Linguistic Feature Analysis ：We employed the Linguistic Inquiry and Word Count (LIWC) framework to quantify psycholinguistic features in the text. The analysis was based on the official LIWC2015 Simplified Chinese dictionary (Pennebaker et al., 2015), which includes validated categories for emotional, cognitive, and social processes. A custom Python script was used to calculate the proportion of words belonging to each category relative to the total word count for each text unit (i.e., participant response or partner response), allowing for standardized comparisons between groups. Topic Modeling ：To uncover the latent thematic structure of the dialogues, we utilized the BERTopic (v0.16.0) model (Grootendorst, 2022). The analysis pipeline consisted of three steps: (1) Text Vectorization: Preprocessed texts were converted into high-dimensional vectors using the SentenceTransformer model ('paraphrase-multilingual-MiniLM-L12-v2'). (2) Topic Generation: The resulting text embeddings were fed into the BERTopic model, which performed topic modeling via clustering and the c-TF-IDF algorithm. Key parameters were set to language='chinese (simplified)', nr_topics='auto', and min_topic_size=10 to automatically determine the optimal number of topics. (3) Topic Interpretation and Naming: To ensure accurate and readable interpretations of the generated topics, we adopted a human-AI collaborative approach (cf. Alharbi et al., 2024). The keyword lists and representative sentences for each topic generated by BERTopic were provided as input to a large language model, ChatGPT (based on the GPT-5 architecture), to generate concise and accurate topic names and descriptions. Statistical Testing To compare language use between the AI and Counsellor groups, we conducted Chi-square tests of independence, followed by standardized residual analysis. These tests were used to determine whether there were significant differences in the relative frequencies of LIWC categories and in the distribution of documents across the identified BERTopic themes. The statistical analysis relied on the pandas (2.3.1), numpy (2.3.2), and SciPy (1.14.1) Python libraries. Result Linguistic Analysis of Role-Playing Interaction Texts To further explore the psychological mechanisms underlying the observed differences in communication intention between the AI and Counsellor groups, we analyzed the dialogue texts generated during the role-playing interactions. Specifically, we examined whether the linguistic styles of the AI and the human counsellor differed, and whether participants, in turn, used language differently when interacting with an AI versus a human partner. Our analysis, conducted using a custom Python script, revealed distinct differences in the Linguistic Inquiry and Word Count (LIWC) category usage between the AI and the counsellor. As illustrated in Figure 5A, which displays the proportions of several key LIWC categories, the largest discrepancies were observed in the categories of Affective Processes, Function Words, Informal Language, and Social Processes. Word usage in other categories was comparatively similar. Subsequent Chi-square tests confirmed that the differences between the AI and the counsellor in these four categories were statistically significant (see Table 3 for details). We then compared the linguistic features of the participants' own texts from the two experimental groups. The results indicated that despite interacting with different partners (AI vs. counsellor), participants themselves showed no significant differences in their use of LIWC word categories (see Figure 5B and Table 4). Topic Modeling of Role-Playing Interaction Texts We further employed BERTopic to conduct topic modeling on the interaction texts, aiming to explore potential differences in the thematic content of the dialogues between the experimental groups. The topic modeling of the partner-generated texts (from the AI and the counsellor) yielded three distinct topics. Figure 6 (left) displays the top five keywords for each topic. Based on these keywords and representative texts, and using the previously described GPT-5 assisted method, the three topics were named as follows (see also Supplementary Table 1): Topic 0: Sincere Apology and Self-Reflection (characterized by keywords related to apology and introspection); Topic 1: Intensifying Regret and Alleviating Tension (marked by expressions of remorse and efforts to de-escalate); and Topic 2: Emotional Affirmation and Relationship Repair (encompassing keywords related to validating feelings and hopes for reconciliation). We then analyzed the distribution of AI- versus counsellor-generated texts across these three topics (see Figure 6A). A Chi-square test revealed a significant difference in this distribution (see Figure 6C for standardized residuals). Specifically, the AI's responses were significantly more likely to fall under Topic 0 (Sincere Apology and Self-Reflection) compared to the counsellor's. Conversely, the counsellor's responses were significantly more concentrated in Topic 2 (Emotional Affirmation and Relationship Repair). The two partners did not differ significantly in their use of Topic 1. The analysis of participant-generated texts yielded four distinct topics, with the top five keywords for each shown in Figure 6 (right). Following the same interpretation procedure (see also Supplementary Table 2), these four topics were named: Topic 0: Emotional Reaction and Privacy Protection (focusing on expressions of feeling and defending personal boundaries); Topic 1: Brief Responses and Emotional Buffering (characterized by short, tension-reducing replies); Topic 2: Accepting Apology and Setting Boundaries (containing keywords related to accepting apologies while establishing future expectations); and Topic 3: Understanding, Acceptance, and Relationship Repair (focusing on reconciliation and moving forward). Finally, we compared the distribution of topics used by participants who interacted with the AI versus those who interacted with the counsellor (see Figure 6B). A Chi-square test revealed no significant difference in the topic distribution between the two groups (see Figure 6D), indicating that participants used similar thematic content regardless of their interaction partner. We employed topic similarity analysis and visualizations of the topic spatial distribution to visually inspect the quality of the topic clusters. For the partner-generated texts, although the topic similarity analysis indicated some overlap between Topic 0 and Topic 1 (see Supplementary Figure 1, left), the visualization of the topic space showed that the clusters were well-differentiated (see Supplementary Figure 2, left). Furthermore, a qualitative review of the semantic content of Topic 0 and Topic 1 confirmed that they represented distinct concepts. Therefore, we retained both topics rather than merging them. Similarly, for the four topics generated from the participant texts, the analyses showed both a degree of thematic similarity and clear differentiation between the clusters (see Supplementary Figure 1, right, and Supplementary Figure 2, right). Discussion2 The results of Study 2a replicated a key finding from Study 1: role-playing, whether with an AI or a human counsellor, led to significant improvements in participants' retaliatory intentions and levels of understanding compared to the control group. This result provides robust support for the generalizability of structured role-playing as an intervention paradigm. It suggests that any medium capable of providing external interaction and feedback—thereby disrupting an individual's egocentric perspective and prompting perspective-taking—will be more effective than mere internal rumination. This further confirms the immense potential of AI to function as a "cognitive scaffold" for practicing conflict resolution skills. Strikingly, however, when directly comparing the intervention effects of the AI against the human expert, we found that while the two were equally effective at reducing retaliation and increasing understanding, the AI was significantly superior to the human expert in the crucial dimension of fostering communication intention. This finding not only supports recent research by Ayers et al. (Ayers et al., 2023)but also offers a counter-intuitive perspective: in certain conflict situations, the "algorithmic empathy" of an AI, which lacks embodied emotional experience, may possess unique advantages. This superiority does not stem from an AI's ability to "feel" in the same way humans do. Rather, it likely originates from its vast underlying data foundation, which enables it to provide efficient, consistent, and unbiased feedback. This allows the AI to offer precise, problem-focused insights and solutions that may be difficult for even a human expert to consistently match(Olawade et al., 2024). To delve into the underlying mechanisms of the AI's superior performance, we conducted an in-depth analysis of the dialogue texts using LIWC and BERTopic. The results revealed the central role of affective responses in both role-playing conditions; both the AI and the human counsellor used a high proportion of emotional words, confirming that affective feedback is an indispensable component of conflict mediation. However, a key difference emerged: the AI's language contained a significantly higher proportion of function words, whereas the human counsellor's language was richer in affective words. This aligns with the theoretical distinction between the two primary components of empathy: affective and cognitive(Deutsch & Madle, 1975). Affective empathy is typically defined as a vicarious, instinctual response to another's emotional state, whereas cognitive empathy involves intentionally taking on another's perspective to infer their thoughts and feelings(Shamay-Tsoory, 2011). Although the AI lacks embodied feelings, it technically recognizes and provides affective feedback to the user's emotional expressions. Crucially, compared to the human counsellor, it places a greater focus on functional responses, offering objective, problem-solving advice and suggestions. We define this specific form of empathy as "Rational Empathy." This focus on providing tangible advice may have sent a potent signal to participants that "communication is a viable pathway to resolving the problem," thereby increasing their motivation to engage in it. Furthermore, the analysis revealed no significant differences in the overall linguistic profiles of the participants themselves across the two groups. This null finding helps to validate the effectiveness of our blind experimental manipulation, suggesting that confounding factors such as "algorithm aversion" did not significantly interfere with the results(Jussupow et al., 2020). When we focus further on the differences in response styles, a clear pattern emerges: the human counsellor's dialogue was more oriented towards "Emotional Affirmation and Relationship Repair." This aligns with the principles of person-centered or humanistic training, which emphasizes the client's deeper feelings and emotional experiences, prioritizing the emotional harm caused over the situational facts of the conflict itself(Rogers, 2012). In contrast, the AI's responses demonstrated a more functional and goal-oriented approach. Its dialogue content was more heavily focused on "Sincere Apology and Self-Reflection" and "Intensifying Regret and Alleviating Tension." The AI's strength lies in its high efficiency, rapidly de-escalating the conflict's intensity and using non-confrontational responses to lower the participant's sense of psychological threat. The human counsellor's strength, conversely, lies in processing the participant's deeper feelings; however, due to the limited length of the conversation and the immediacy of the outcome measurement, the benefits of this deeper approach may not have been fully captured in our results. The findings from Study 2b provide robust theoretical support for the AI's superiority in fostering communication intention. The AI's "Rational Empathy", through its efficient and non-confrontational replies, rapidly de-escalated the conflict and sent a potent underlying signal to the participant: "communication is a fast and effective way to solve this problem." This problem-focused and action-oriented style may be particularly effective at motivating communication within the context of an immediate conflict. In contrast, while a human counsellor can process deep-seated emotions, the impact of this approach may have a lagged effect that is not fully manifested in a brief, time-limited interaction. General discussion This research aimed to explore the potential of AI-enabled role-playing (AI-RP) in managing interpersonal conflicts and to compare its effectiveness with traditional human-led role-playing. Across two experimental studies, we not only validated the feasibility of AI-RP as an effective intervention tool but also uncovered its unique advantages and the underlying mechanisms driving them. Our findings indicate that structured role-playing is a potent method for breaking ingrained patterns of conflict. Study 1 demonstrated that AI-RP significantly improved participants' interpersonal conflict resolution strategies, specifically by enhancing understanding and communication while reducing retaliatory intentions. Study 2a further consolidated this finding, showing that role-playing with either an AI or a human expert was superior to solitary rumination. This provides strong evidence that any form of structured role-playing intervention, by offering an external and continuously interactive partner, can effectively disrupt the egocentric thinking that is habitual in conflicts. This interactive process compels a "perspective-taking" mechanism, and this forced cognitive reappraisal is core to prompting more constructive responses. It is crucial to note, however, that this positive effect did not generalize to a new, parallel context. This suggests that AI-RP, in its current form, may function as a context-dependent "cognitive scaffold." Future applications must explore how to help users internalize these learned skills to handle real-world, high-load conflict situations. Another key finding is the unique advantage of AI-RP in fostering participants' willingness to communicate. Study 2a revealed that the AI was even more effective than a human expert at enhancing communication intentions. Our text analysis uncovered the mechanism behind this counter-intuitive finding. The AI's superiority stems not from superior emotionality but from its unique combination of "Rational Empathy" and a problem-focused approach. Although the AI also used a high proportion of affective words, its dialogue patterns were more efficiency-oriented. By rapidly acknowledging fault and offering solutions, it sent an implicit signal that "communication can solve this problem." This approach effectively lowered the participants' sense of psychological threat, making them more willing to initiate communication. This stands in stark contrast to the human expert's style, which, being more focused on emotional processing, may have a less immediate impact. Crucially, the findings of this research are not intended to suggest that AI will fully replace human counselors but rather to highlight the value of their complementary coexistence. The topic analysis in Study 2b further substantiates this point. The human counselor's dialogue was more oriented towards "Emotional Affirmation and Relationship Repair," which aligns with the deep emotional processing emphasized in humanistic therapies, but its full effects may require a longer timeframe to manifest(Howard et al., 1986). In contrast, the AI's advantage lies in its efficient, non-confrontational handling of immediate conflicts, rapidly de-escalating their intensity and creating favorable conditions for problem-solving. In future practice, AI could serve as an initial intervention tool to help individuals quickly de-escalate a conflict and build the willingness to communicate. Human experts could then build upon this foundation to provide deeper emotional support and relational repair. This human-AI collaborative model, rather than a simple substitution of humans by AI, could provide more comprehensive, efficient, and scalable solutions for resolving interpersonal conflict(Devillers, 2021). This research provides a solid empirical foundation for AI-assisted psychological interventions and interpersonal skills training. We have demonstrated that in dynamic role-playing, an AI can act as a novel "cognitive scaffold" to activate users' perspective-taking in conflicts. Furthermore, we have re-examined the concept of AI "empathy" by proposing the construct of "Rational Empathy," offering a new direction for the design of AI in the field of Human-Computer Interaction (HCI). AI-RP can be developed into various applications, such as a simulation training tool for workplace mediation, a practice platform for family communication, or an educational tool for developing conflict resolution skills in adolescents. Its efficiency, accessibility, and scalability make large-scale, personalized psychological support a tangible possibility. Nevertheless, this study has several limitations that should be acknowledged. The lack of transferability of the observed AI-RP effects suggests that future research should explore how longer-term training might enhance skill internalization. Second, our assessment was based on immediate behavioral intentions and did not capture the long-term effects of the intervention. Additionally, the individual characteristics of the single human expert may limit the generalizability of our conclusions. Finally, cross-cultural factors, such as the emphasis on harmony in Chinese culture(Leung et al., 2011), and their potential impact on the results warrant further investigation. Future research could also explore the development of multimodal AI-RP systems that incorporate voice and facial expression recognition to enhance realism(Li & Deng, 2020). Moreover, investigating how to tailor the AI's intervention strategies based on users' personality traits presents an exciting avenue for future inquiry. Declarations Ethics approval and consent to participate Our research is conducted in strict accordance with the ethical principles outlined in the Declaration of Helsinki, ensuring the utmost protection of the safety, privacy, and rights of all participants. The studies involving humans were approved by the Ethics Committee of Anhui University of Science and Technology. The studies were conducted in accordance with the local legislation and institutional requirements. The entire process adhered to the principle of voluntary participation. Participants were free to choose to withdraw from the experiment. After the study was completed, the investigator clearly explained the purpose of the study to all participants. Availability of data and materials All data generated or analyzed during this study are included in this published article. Funding This work was supported by the University-Level Quality Engineering Project of Anhui Science and Technology University (Project No. Xj2024200). Authors' contributions TTY: Conceptualization, Formal analysis, Investigation, Writing – original draft, Funding acquisition. HLM: Conceptualization, Methodology. HYW: Supervision. HDS: Conceptualization, Project administration, Methodology, Supervision, Writing – original draft, Writing – review & editing. Acknowledgments We sincerely acknowledge all participants involved in this research and those who assisted in the recruitment process. This study was supported by the University-Level Quality Engineering Project of Anhui University of Science and Technology (Project No. Xj2024200). References Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., & Anadkat, S. (2023). Gpt-4 technical report. arXiv preprint arXiv:2303.08774 . Adam, M., Wessel, M., & Benlian, A. (2021). Ai-based chatbots in customer service and their effects on user compliance. Electronic Markets , 31 (2), 427-445. https://doi.org/10.1007/s12525-020-00414-7 Adamopoulou, E., & Moussiades, L. (2020). An overview of chatbot technology. IFIP international conference on artificial intelligence applications and innovations, Aydoğan, R., Baarslag, T., & Gerding, E. (2021). Artificial intelligence techniques for conflict resolution. Group Decision and Negotiation , 30 (4), 879-883. https://doi.org/10.1007/s10726-021-09738-x Ayers, J. W., Poliak, A., Dredze, M., Leas, E. C., Zhu, Z., Kelley, J. B., Faix, D. J., Goodman, A. M., Longhurst, C. A., Hogarth, M., & Smith, D. M. (2023). Comparing physician and artificial intelligence chatbot responses to patient questions posted to a public social media forum. JAMA Internal Medicine , 183 (6), 589-596. https://doi.org/10.1001/jamainternmed.2023.1838 Chaturvedi, A. (2024). Exploring empathy in artificial intelligence: Synthesis and paths for future research. Information Discovery and Delivery , 53 (3), 406-424. https://doi.org/10.1108/IDD-03-2024-0048 Davis, M. H. (1983). Measuring individual differences in empathy: Evidence for a multidimensional approach. Journal of Personality and Social Psychology , 44 (1), 113. https://doi.org/10.1037//0022-3514.44.1.113 Deutsch, F., & Madle, R. A. (1975). Empathy: Historic and current conceptualizations, measurement, and a cognitive theoretical perspective. Human development , 18 (4), 267-287. https://doi.org/10.1159/000271488 Devillers, L. (2021). Human–robot interactions and affective computing: The ethical implications. In Robotics, ai, and humanity: Science, ethics, and policy (pp. 205-211). Springer International Publishing Cham. Feldkamp, T., Langer, M., Wies, L., & König, C. J. (2023). Justice, trust, and moral judgements when personnel selection is supported by algorithms. European Journal of Work and Organizational Psychology , 33 (2), 130-145. https://doi.org/10.1080/1359432x.2023.2169140 Figley, C. R. (2002). Compassion fatigue: Psychotherapists' chronic lack of self care. Journal of clinical psychology , 58 (11), 1433-1441. https://doi.org/10.1002/jclp.10090 Fredrickson, B. L. (2001). The role of positive emotions in positive psychology: The broaden-and-build theory of positive emotions. American Psychologist , 56 (3), 218-226. https://doi.org/10.1037/0003-066X.56.3.218 Galinsky, A. D., Ku, G., & Wang, C. S. (2005). Perspective-taking and self-other overlap: Fostering social bonds and facilitating social coordination. Group processes & intergroup relations , 8 (2), 109-124. https://doi.org/10.1177/1368430205051060 Galinsky, A. D., & Moskowitz, G. B. (2000). Perspective-taking: Decreasing stereotype expression, stereotype accessibility, and in-group favoritism. Journal of Personality and Social Psychology , 78 (4), 708-724. https://doi.org/10.1037/0022-3514.78.4.708 Harel, D. (2004). Computers ltd: What they really can't do . OUP Oxford. Howard, K. I., Kopta, S. M., Krause, M. S., & Orlinsky, D. E. (1986). The dose–effect relationship in psychotherapy. American Psychologist , 41 (2), 159. https://doi.org/10.1037/0003-066X.41.2.159 Hsu, A., & Chaudhary, D. (2023). Ai4pcr: Artificial intelligence for practicing conflict resolution. Computers in Human Behavior: Artificial Humans , 1 (1), 100002. https://doi.org/https://doi.org/10.1016/j.chbah.2023.100002 Jia, F. (2014). The impact of heterogeneous high self-esteem on college students' interpersonal conflict resolution style under situational forgiveness NanJing Normal University]. Jussupow, E., Benbasat, I., & Heinzl, A. (2020). Why are we averse towards algorithms? A comprehensive literature review on algorithm aversion. Kalla, D., Smith, N., Samaah, F., & Kuraku, S. (2023). Study and analysis of chat gpt and its impact on different fields of study. International journal of innovative science and research technology , 8 (3). Kim, T., & Peng, W. (2024). Do we want ai judges? The acceptance of ai judges’ judicial decision-making on moral foundations. AI & SOCIETY , 40 (5), 3683-3696. https://doi.org/10.1007/s00146-024-02121-9 Klimecki, O. M. (2019). The role of empathy and compassion in conflict resolution. Emotion Review , 11 (4), 310-325. https://doi.org/10.1177/1754073919838609 Leung, K., Brew, F. P., Zhang, Z.-X., & Zhang, Y. (2011). Harmony and conflict: A cross-cultural investigation in china and australia. Journal of Cross-Cultural Psychology , 42 (5), 795-816. https://doi.org/10.1177/0022022110363474 Levy, K. N., & Scala, J. W. (2012). Transference, transference interpretations, and transference-focused psychotherapies. Psychotherapy , 49 (3), 391-403. https://doi.org/10.1037/a0029371 Li, S., & Deng, W. (2020). Deep facial expression recognition: A survey. IEEE transactions on affective computing , 13 (3), 1195-1215. https://doi.org/10.1109/TAFFC.2020.2981446 Lutgen-Sandvik, P., & Sypher, B. D. (2009). Destructive organizational communication. Processes, Consequences and Constructive Ways of Organizing . Mead, G. H. (2015). Mind, self & society . University of Chicago Press. Olawade, D. B., Wada, O. Z., Odetayo, A., David-Olawade, A. C., Asaolu, F., & Eberhardt, J. (2024). Enhancing mental health with artificial intelligence: Current trends and future prospects. Journal of Medicine, Surgery, and Public Health , 3 , 100099. https://doi.org/10.1016/j.glmedi.2024.100099 Pagani, A. F., Donato, S., Parise, M., Bertoni, A., Iafrate, R., & Schoebi, D. (2019). Explicit stress communication facilitates perceived responsiveness in dyadic coping. Frontiers in psychology , 10 , 401. https://doi.org/10.3389/fpsyg.2019.00401 Rogers, C. (2012). Client centered therapy (new ed) . Hachette UK. Rosler, N., Cohen-Chen, S., & Halperin, E. (2015). The distinctive effects of empathy and hope in intractable conflicts. Sage Journals . https://doi.org/10.1177/0022002715569772 Shamay-Tsoory, S. G. (2011). The neural bases for empathy. The Neuroscientist , 17 (1), 18-24. https://doi.org/10.1177/1073858410379268 Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive Science , 12 (2), 257-285. https://doi.org/10.1016/0364-0213(88)90023-7 Taguchi, N. (2023). Using immersive virtual reality for the assessment of intercultural conflict mediation. Language Learning & Technology , 27 (3), 108-128. https://doi.org/10.64152/10125/73534 Thoits, P. A. (2011). Mechanisms linking social ties and support to physical and mental health. Journal of health and social behavior , 52 (2), 145-161. https://doi.org/10.1177/0022146510395592 Vygotsky, L. S., & Cole, M. (1978). Mind in society: Development of higher psychological processes . Harvard University Press. Additional Declarations The authors declare no competing interests. Supplementary Files SupplementaryFigureandTable.pdf Supplementary Table 1. Topic labeling and representative texts of AI and counselor responses (BERTopic + ChatGPT). Supplementary Table 2. Topic labeling and representative texts of participant–AI and participant–counselor interactions (BERTopic + ChatGPT). Supplementary Figure 1. Similarity among topics in AI and counselor response texts (left) and similarity between participant–AI and participant–counselor interaction texts (right). Supplementary Figure 2. Spatial distribution of topics in AI and counselor response texts (left) and in participant–AI and participant–counselor interaction texts (right). 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-7699927","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":519761214,"identity":"db95333e-9ccb-44c2-ab84-7bdd49dbd185","order_by":0,"name":"Tongtong Ye","email":"","orcid":"","institution":"college of Humanities, Anhui Science and Technology University, ChuZhou, China","correspondingAuthor":false,"prefix":"","firstName":"Tongtong","middleName":"","lastName":"Ye","suffix":""},{"id":519761215,"identity":"c4ef7bf2-b4a0-4694-ad34-57b6c0e32b50","order_by":1,"name":"Haoliang Mai","email":"","orcid":"","institution":"Faculty of Educational Sciences and Technology, Universiti Teknologi Malaysia, Johor Bahru, Malaysia","correspondingAuthor":false,"prefix":"","firstName":"Haoliang","middleName":"","lastName":"Mai","suffix":""},{"id":519761216,"identity":"1178a16b-f86c-414f-aa47-ce30720d9a81","order_by":2,"name":"Hongyu Wang","email":"","orcid":"","institution":"college of Humanities, Anhui Science and Technology University, ChuZhou, China","correspondingAuthor":false,"prefix":"","firstName":"Hongyu","middleName":"","lastName":"Wang","suffix":""},{"id":519761217,"identity":"23b5b42f-3793-4a89-8e0d-f0e912049de5","order_by":3,"name":"Haodong Su","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABCElEQVRIiWNgGAWjYFACHiBmA2Jm5oOP/1TIgcUOPCBKCztbsgHPGWOIlgSitPDzmEnwtkG0MODTIj8j9+CHD2U2efLOQC2S8wzkDK4dfgi0xU5OtwG7FoMbecmSM86lFRseZiu2MNxmYCw5O80AqCXZ2OwADi0SOQbSvG2HEzc2M2+8kbjtT2K/dAJIy4HEbTi0yM/IMf4N0QLUfnCOQWKbdPoHvFoYbuSYgW2Zz8xiJNnYYAC0JQe/LQZn3phZAv2SuIGZLdmY4RjILzkFBxIMcPtFvj3H+AYwxBLn9x8++JihBhhit9M3f/hQYSeHSwvCOlQFBgSUg61rIELRKBgFo2AUjEwAAFtwYDkkzwuQAAAAAElFTkSuQmCC","orcid":"","institution":"college of Humanities, Anhui Science and Technology University, ChuZhou, China","correspondingAuthor":true,"prefix":"","firstName":"Haodong","middleName":"","lastName":"Su","suffix":""}],"badges":[],"createdAt":"2025-09-24 06:15:26","currentVersionCode":1,"declarations":{"humanSubjects":true,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":true,"humanSubjectConsent":true,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7699927/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7699927/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":92233866,"identity":"9443ac92-d0c0-460f-8e3d-deab9f8b1814","added_by":"auto","created_at":"2025-09-26 07:00:36","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":59995,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-7699927/v1/e3e6f41ab0b6da992761683c.docx"},{"id":92233867,"identity":"1668352c-e353-45ba-b446-653c77652f42","added_by":"auto","created_at":"2025-09-26 07:00:37","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":342,"visible":true,"origin":"","legend":"","description":"","filename":"rs7699927.json","url":"https://assets-eu.researchsquare.com/files/rs-7699927/v1/0e49fdb65797381283fbb028.json"},{"id":92234697,"identity":"33e25fb1-25ad-4e6a-8d78-3fc99100e475","added_by":"auto","created_at":"2025-09-26 07:08:37","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":105491,"visible":true,"origin":"","legend":"","description":"","filename":"rs76999270enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7699927/v1/74fb99e48f31900328a32f02.xml"},{"id":92233876,"identity":"4bb00069-ca7e-409b-bcbe-11f250ebe40c","added_by":"auto","created_at":"2025-09-26 07:00:37","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":99568,"visible":true,"origin":"","legend":"","description":"","filename":"rs76999270structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7699927/v1/8133814c912d3ad6466709dd.xml"},{"id":92233878,"identity":"df69a1c6-91fe-4b75-89a0-02877d16fc22","added_by":"auto","created_at":"2025-09-26 07:00:37","extension":"html","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":112092,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7699927/v1/f3fb133826f1ee17496c1925.html"},{"id":92234696,"identity":"16feb399-5bf3-4efe-812c-282c41cafd6a","added_by":"auto","created_at":"2025-09-26 07:08:37","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":47919,"visible":true,"origin":"","legend":"\u003cp\u003eExperimental procedure of Study 1\u003c/p\u003e","description":"","filename":"fig.1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7699927/v1/08388da45e42c4201ce31b6a.jpg"},{"id":92234695,"identity":"d2c13cd1-9b89-413f-9035-66b490cfbd33","added_by":"auto","created_at":"2025-09-26 07:08:37","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":31597,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThe AI positive response group exhibited significantly higher willingness to communicate, but did not show lower willingness to seek revenge or higher willingness to understand.\u003cbr\u003e\n\u003c/strong\u003e(A) No significant difference was observed in revenge between the positive response (PR) and negative response (NR) groups, but both scored significantly lower than the control group (CON) (One-way ANOVA, between-group effect: F (2, 76) = 3.76, \u003cem\u003ep\u003c/em\u003e = 0.028, ω² = 0.07; Tukey post-hoc comparisons: PR vs. NR: \u003cem\u003ep\u003c/em\u003e = 0.615; PR vs. CON: \u003cem\u003ep\u003c/em\u003e = 0.024; NR vs. CON: \u003cem\u003ep\u003c/em\u003e = 0.144). (B) Both positive and negative responses significantly increased participants’ willingness to understand (One-way ANOVA, between-group effect: F (2, 76) = 7.17, \u003cem\u003ep\u003c/em\u003e = 0.001, ω² = 0.135; Tukey post-hoc comparisons: PR vs. NR: \u003cem\u003ep\u003c/em\u003e= 0.941; PR vs. CON: \u003cem\u003ep \u003c/em\u003e= 0.003; NR vs. CON: \u003cem\u003ep\u003c/em\u003e = 0.005). (C) Positive AI responses significantly enhanced participants’ willingness to communicate (One-way ANOVA, between-group effect: F (2, 76) = 7.31, \u003cem\u003ep\u003c/em\u003e = 0.001, ω² = 0.138; Tukey post-hoc comparisons: PR vs. NR: \u003cem\u003ep\u003c/em\u003e = 0.004; PR vs. CON: \u003cem\u003ep \u003c/em\u003e= 0.003; NR vs. CON: p = 0.938). (D, E, F) provide detailed distributions across groups for \u003cstrong\u003eRevenge, Understanding, \u003c/strong\u003eand\u003cstrong\u003e Communication,\u003c/strong\u003e respectively.\u003c/p\u003e","description":"","filename":"fig.2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7699927/v1/dcc2718e1edaf1faf5433d28.jpg"},{"id":92233872,"identity":"0403ace4-ec99-4801-8673-0878697eee3d","added_by":"auto","created_at":"2025-09-26 07:00:37","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":45696,"visible":true,"origin":"","legend":"\u003cp\u003eExperimental procedure of Study 2\u003c/p\u003e","description":"","filename":"fig.3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7699927/v1/3d723e5858a03ba4f1461ba3.jpg"},{"id":92233879,"identity":"3f0ed2a0-d100-4fa2-889a-6d4f353341e9","added_by":"auto","created_at":"2025-09-26 07:00:37","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":33227,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eRole-playing with the AI, rather than with a counselor, significantly increased participants’ willingness to communicate, while no significant differences were observed in willingness to seek revenge or willingness to understand.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(A) Interaction with either a counselor or the AI reduced participants’ willingness to seek revenge (One-way ANOVA, between-group effect: F (2, 80) = 7.267, \u003c/strong\u003e\u003cem\u003e\u003cstrong\u003ep \u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003e= 0.0013; Tukey post-hoc comparisons: Control vs. Counselor: \u003c/strong\u003e\u003cem\u003e\u003cstrong\u003ep\u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003e = 0.005; Control vs. AI: \u003c/strong\u003e\u003cem\u003e\u003cstrong\u003ep\u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003e = 0.0034; Counselor vs. AI: \u003c/strong\u003e\u003cem\u003e\u003cstrong\u003ep \u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003e= 0.9750).\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(B) Interaction with either a counselor or the AI increased participants’ willingness to understand (One-way ANOVA, between-group effect: F (2, 80) = 8.251, \u003c/strong\u003e\u003cem\u003e\u003cstrong\u003ep \u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003e= 0.0006; Tukey post-hoc comparisons: Control vs. Counselor: \u003c/strong\u003e\u003cem\u003e\u003cstrong\u003ep\u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003e = 0.0017; Control vs. AI: \u003c/strong\u003e\u003cem\u003e\u003cstrong\u003ep \u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003e= 0.0026; Counselor vs. AI: \u003c/strong\u003e\u003cem\u003e\u003cstrong\u003ep \u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003e= 0.9995).\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(C) Compared with the counselor, the AI significantly enhanced participants’ willingness to communicate (One-way ANOVA, between-group effect: F (2, 80) = 8.243, \u003c/strong\u003e\u003cem\u003e\u003cstrong\u003ep \u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003e= 0.0006; Tukey post-hoc comparisons: Control vs. Counselor: \u003c/strong\u003e\u003cem\u003e\u003cstrong\u003ep \u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003e= 0.4056; Control vs. AI: \u003c/strong\u003e\u003cem\u003e\u003cstrong\u003ep \u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003e= 0.004; Counselor vs. AI: \u003c/strong\u003e\u003cem\u003e\u003cstrong\u003ep \u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003e= 0.0198).\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(D) Detailed proportions of conflict outcomes following interactions with the AI and the counselor, categorized as conflict resolved, partially resolved, or unresolved.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"fig.4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7699927/v1/ba549f49da469cd1c2b5963a.jpg"},{"id":92233870,"identity":"2b7cd96c-fe94-46d1-88d8-e31316c58fb4","added_by":"auto","created_at":"2025-09-26 07:00:37","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":21890,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eWord category analysis of using the official Chinese LIWC dictionary.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA. Word category analysis of counselor and AI responses to participants’ texts\u003c/p\u003e\n\u003cp\u003eB. Word category analysis of participant–counselor and participant–AI interaction texts\u003c/p\u003e","description":"","filename":"fig.5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7699927/v1/4347e8ac55e1c721489564b1.jpg"},{"id":92233875,"identity":"a869f96c-65e9-431a-877a-8b7597009b45","added_by":"auto","created_at":"2025-09-26 07:00:37","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":39254,"visible":true,"origin":"","legend":"\u003cp\u003eTopic analysis of AI or counselor responses (left) and participant responses (right) using BERTopic.\u003c/p\u003e","description":"","filename":"fig.6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7699927/v1/bfa8e1d5b420f33d50be3086.jpg"},{"id":92233877,"identity":"f8e51f29-4dfc-4d95-982d-e340896f7ce9","added_by":"auto","created_at":"2025-09-26 07:00:37","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":36041,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAI responses exhibited significantly different topic characteristics compared with counselor responses, whereas participants’ responses to the two showed no significant differences in topic distribution.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(A) Proportional distribution of AI and counselor responses across Topic_0, Topic_1, and Topic_2.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(B) Proportional distribution of participants’ responses across Topic_0, Topic_1, Topic_2, and Topic_3.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(C) Distribution of topics across groups (AI vs. counselor). A chi-square test indicated a significant association between group and topic, χ² (2, N = 316) = 53.50, \u003c/strong\u003e\u003cem\u003e\u003cstrong\u003ep\u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003e \u0026lt; 0.001. Standardized residuals showed that AI responses were more likely to be categorized as Topic_0, whereas counselor responses were more likely to be categorized as Topic_2. No significant differences were found for Topic_1.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(D) Distribution of topics in participants’ responses to AI and counselor. The chi-square test indicated no significant association between group and topic, χ² (3, N = 291) = 3.77, \u003c/strong\u003e\u003cem\u003e\u003cstrong\u003ep \u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003e= 0.287.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"fig.7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7699927/v1/5a2e40f6bbbe8993059d893f.jpg"},{"id":92235447,"identity":"ec7bd4ed-7ca7-4dd6-a7da-30ab15458532","added_by":"auto","created_at":"2025-09-26 07:16:38","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1896041,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7699927/v1/8701f65b-cf1e-4661-b01b-79b1ba6b3d16.pdf"},{"id":92233868,"identity":"d8e141c7-d987-482a-99bb-891c6d0ed89b","added_by":"auto","created_at":"2025-09-26 07:00:37","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":426981,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary Table 1. Topic labeling and representative texts of AI and counselor responses (BERTopic + ChatGPT).\u003c/p\u003e\n\u003cp\u003eSupplementary Table 2. Topic labeling and representative texts of participant–AI and participant–counselor interactions (BERTopic + ChatGPT).\u003c/p\u003e\n\u003cp\u003eSupplementary Figure 1. Similarity among topics in AI and counselor response texts (left) and similarity between participant–AI and participant–counselor interaction texts (right).\u003c/p\u003e\n\u003cp\u003eSupplementary Figure 2. Spatial distribution of topics in AI and counselor response texts (left) and in participant–AI and participant–counselor interaction texts (right).\u003c/p\u003e","description":"","filename":"SupplementaryFigureandTable.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7699927/v1/c3d5422809572ea0ad0c4967.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eAI's Rational Empathy Promotes Reconciliation in Conflict: Evidence from Behavioral Experiments, Linguistic Analysis, and Topic Modeling\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eWhile the use of chatbots for obtaining basic, factual information has long been commonplace(Adamopoulou \u0026amp; Moussiades, 2020), the advent of Large Language Models (LLMs) since 2022, spearheaded by OpenAI's ChatGPT, has profoundly reshaped the paradigm of human-chatbot interaction(Kalla et al., 2023). Diverging from the utilitarian expectations placed on traditional virtual assistants, user expectations for LLM-based chatbots are increasingly shifting towards the fulfillment of emotional and psychological needs(Achiam et al., 2023). This shift highlights the immense potential of LLMs as tools for fostering human psychological well-being and social competencies. Among the myriad of human social experiences, interpersonal conflict represents a particularly pervasive and challenging domain. The effective management of such conflicts is crucial for maintaining mental health, fostering interpersonal harmony, and ensuring social stability(Thoits, 2011). LLMs are now being applied as mediators to address real-world interpersonal conflicts in areas such as automated negotiation, group decision-making, argumentation, preference aggregation, and human-computer interaction(Aydoğan et al., 2021). This application persists despite the well-documented phenomenon of \"algorithm aversion\"(Feldkamp et al., 2023; Kim \u0026amp; Peng, 2024), which encapsulates public skepticism about whether an impersonal AI can genuinely understand and navigate complex human emotions. Such skepticism is particularly pronounced in domains heavily reliant on empathy. Nevertheless, the use of LLMs as auxiliary tools for emotional support and conflict mediation is becoming increasingly common(Hsu \u0026amp; Chaudhary, 2023).\u003c/p\u003e\n\u003cp\u003eAt the core of effective interpersonal conflict mediation lies empathy—the ability to experience the world from another's perspective(Chaturvedi, 2024). In conflict situations, empathy can reduce aggressive attitudes(Rosler et al., 2015)\u0026nbsp;and lead to lower levels of escalation(Klimecki, 2019). Through extensive training on vast datasets, LLMs enable computers to express empathy by analyzing and responding to user inputs(Adam et al., 2021). However, translating an LLM's capacity for empathy into an effective conflict intervention requires a suitable medium of implementation. Drawing on symbolic interactionism, Mead(Mead, 2015)proposed that the self develops through the continuous interaction between the \"I\" and the \"Me,\" and that we come to understand others' perspectives and attitudes through role-taking. Reinforcing this, extensive research in social psychology has confirmed that role-playing is a powerful intervention technique. By immersing individuals in simulated conflict scenarios, it effectively promotes perspective-taking, elicits cognitive and emotional responses, and ultimately improves their conflict resolution strategies(Lutgen-Sandvik \u0026amp; Sypher, 2009; Taguchi, 2023). Role-playing thus offers an ideal research paradigm for the present study: it serves as both an established technique for modifying conflict responses and a controlled, safe environment for precisely comparing the mediating effects of an LLM acting as a third party.\u003c/p\u003e\n\u003cp\u003ePrevious research has established that individuals typically seek external emotional and social support when facing interpersonal conflicts or stressful events(Thoits, 2011). Positive and empathic responses not only buffer negative emotions but also facilitate constructive communication and conflict resolution(Pagani et al., 2019). Lutgen-Sandvik and Sypher(Lutgen-Sandvik \u0026amp; Sypher, 2009)\u0026nbsp;categorized conflict coping styles into two types: destructive approaches, such as blaming, attacking, and retaliating, and constructive approaches, which include understanding, perspective-taking, and constructive communication. The present study is the first to systematically investigate the effectiveness of AI-enabled role-playing (AI-RP) in influencing participants' interpersonal conflict resolution strategies through a behavioral experiment. Accordingly, we propose our first hypothesis (H1): Participants who engage in role-playing with an AI will adopt more constructive conflict resolution strategies than those in a no-interaction control group. Specifically, participants interacting with an AI exhibiting a positive, empathic response style will demonstrate significantly more constructive approaches (i.e., higher intentions for understanding and communication, and lower intention for retaliation) compared to both participants interacting with a negative-response AI and those in the control group.\u003c/p\u003e\n\u003cp\u003eA deeper question then arises: when both an AI and a professionally trained human counselor provide empathy during role-playing, which approach is more effective in improving conflict resolution outcomes? Empathy is a multidimensional construct, comprising both affective and cognitive dimensions(Davis, 1983). The empathy of a human counselor is rooted in genuine emotional experience, a significant strength that can also be a limitation. For instance, phenomena such as transference and countertransference may lead to biased judgments, while sustained emotional involvement can result in compassion fatigue(Figley, 2002; Levy \u0026amp; Scala, 2012). In contrast, the empathy exhibited by an AI is what we term \"Rational Empathy\"—a form of empathy derived from data-driven pattern recognition and language generation. Devoid of genuine emotional experience, it is immune to emotional fluctuations, biases, or fatigue, allowing it to provide more stable, consistent, and controllable empathic responses(Adam et al., 2021). This sets up a key theoretical tension: the embodied, authentic emotional empathy of a human counselor versus the disembodied, computational \"Rational Empathy\" of an AI. On one hand, a genuine emotional connection may be the critical catalyst for user change. On the other hand, stable and unbiased rational support might offer a safer conversational space in highly emotional conflicts. Existing theories do not offer a clear prediction as to which approach is superior.\u003c/p\u003e\n\u003cp\u003eTherefore, the second core exploratory research question of this study is as follows: In the context of conflict role-playing, which interaction—with an AI or with a human counselor—is more effective at promoting an individual's constructive behaviors in subsequent conflict situations? We will address this question by experimentally comparing the effectiveness of these two intervention modalities. Furthermore, we will conduct an in-depth analysis of the dialogue texts to investigate the potential mechanistic differences between the AI's \"Rational Empathy\" and the human counselor's \"Emotional Empathy.\"\u003c/p\u003e\n\u003cp\u003eTo systematically address the aforementioned research questions and test our hypotheses, we designed a series of two studies. Study 1, a behavioral experiment, was designed to investigate the impact of different AI response styles (positive, negative, and a no-interaction control) during role-playing on participants' subsequent conflict behaviors. Study 2 directly compared positive role-playing interactions with an AI versus a human counselor to explore the differential effects of \"Rational Empathy\" and \"Emotional Empathy\" on conflict resolution. Furthermore, to delve into the underlying psychological mechanisms, we employed a multi-method approach, combining behavioral data with LIWC-based linguistic analysis and BERTopic-based topic modeling. This approach allows us to reveal the multifaceted impacts of different interaction modalities on individuals' cognition, emotions, and behaviors.\u003c/p\u003e"},{"header":"Study 1","content":"\u003cp\u003e\u003cstrong\u003eMethod\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eParticipants\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAn a priori power analysis was conducted using G*Power (Version 3.1) to determine the required sample size. The analysis was set for a MANOVA: Global effects, with parameters of α = 0.05, power (1-β) = 0.80, and a medium effect size (f²(V) = 0.08). The results indicated a minimum required sample of 90 participants. We initially recruited 99 university students. Twenty participants were excluded as they failed to complete the entire experimental procedure, resulting in a final sample of 79 individuals. Participants were randomly assigned to one of three groups: a positive-response group (PR; n = 24, 8 males, 16 females; \u003cem\u003eM\u003c/em\u003e = 19.08, SD = 0.78), a negative-response group (NR; n = 31, 18 males, 13 females; \u003cem\u003eM\u003c/em\u003e= 19.26, SD = 1.06), or a no-interaction control group (CON; n = 24, 9 males, 15 females; \u003cem\u003eM\u003c/em\u003e = 20.63, SD = 1.93). During the experiment, participants were blind to their assigned condition and the study's true hypotheses. This research was approved by the Institutional Review Board of Anhui University of Science and Technology. All participants provided written informed consent prior to the experiment and were informed of their right to withdraw at any time.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eDesign\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA single-factor, between-subjects design was employed. The independent variable was the AI's response style during role-playing, with three levels: positive, negative, and a no-interaction control. The dependent variables were participants' intentions of retaliation, understanding, and communication. This study aimed to examine the effect of the AI's response style on participants' reactions in conflict scenarios and to test the generalization of this effect. To assess generalization, participants' responses were measured in two contexts: (1) a repeated context, identical to the one used in the role-play, and (2) a parallel context, which was thematically similar but distinct from the role-play scenario.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eMaterials and Measures\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eConflict Scenarios\u003c/em\u003e. The scenario for the role-play (and repeated context measure) described a conflict regarding reciprocal responsibilities within a three-person study group. The protagonist had fulfilled their duty of saving a library seat with a power outlet for the group on three occasions, whereas a partner had failed to do so on two of their designated days without prior notice. This negligence resulted in the protagonist being unable to secure an adequate seat for a group discussion, ultimately disrupting their preparation when their device's battery depleted. The parallel context described a conflict in a course group project: the protagonist, as the group leader, publicly pointed out that a member (A) had repeatedly failed to complete their assigned tasks. Subsequently, Member A quit the group chat, locked the shared files, and submitted their own work independently, thereby hindering the group's progress.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eDependent Measures\u003c/em\u003e. Participants' behavioral intentions were measured using four items adapted from prior research on conflict resolution(Jia, 2014). All items were rated on a 5-point Likert scale (1 = Strongly Disagree, 5 = Strongly Agree). Two items assessed retaliation intention (e.g., \"Find a similar opportunity to make him/her feel the same way,\" \"Spread what he/she did to other classmates\"). One item assessed understanding (e.g., \"He/she was just acting according to his/her personality; I can understand him/her\"). One item assessed communication intention (e.g., \"Find a suitable opportunity to have an effective conversation with him/her\").\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eCommunication and AI Tools\u003c/em\u003e. The interaction was conducted via the instant messaging software QQ. To standardize the interaction and control the AI's persona, we employed a \"Wizard of Oz\" technique. A research assistant, unknown to the participant, acted as an intermediary, relaying messages between the participant and the AI model. This setup also allowed us to control the timing and pace of the AI's responses. Following a pilot evaluation of several Large Language Models (LLMs), \"Doubao\" was selected as the AI tool for its superior performance in adhering to the complex instructional prompts required for the positive and negative response conditions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eProcedure\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe experiment was conducted in a standardized laboratory setting over three consecutive days, with conditions counterbalanced across time slots. After providing informed consent, participants were randomly assigned to one of the three conditions (see Figure 1 for a detailed flowchart).\u003c/p\u003e\n\u003cp\u003eParticipants were instructed to begin a conversation via QQ with a \"partner\" (the AI) about the library seat-saving conflict scenario. They were asked to send the first message, after which the interaction proceeded for five turns per participant (totaling 10 messages). In the positive-response (PR) condition, the AI was prompted to provide empathic and understanding responses. In the negative-response (NR) condition, the AI was prompted to be negative and unsupportive. Participants in the control (CON) group did not interact with the AI; instead, after reading the scenario, they were instructed to imagine being in the situation for three minutes.\u003c/p\u003e\n\u003cp\u003eImmediately following the interaction (or imagination task), participants completed a manipulation check, rating their engagement in the role-play on a 5-point scale. Subsequently, they completed an online questionnaire containing the dependent measures for both the repeated and parallel contexts. Finally, all participants were fully debriefed regarding the true purpose of the study and the use of AI.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eData Analysis\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA one-way multivariate analysis of variance (MANOVA) was conducted using JASP (Version 0.95) to examine the overall effect of the AI's response style on the three dependent variables. Pillai's Trace was selected as the test statistic due to its robustness. If the MANOVA yielded a significant result, a series of one-way ANOVAs were performed on each dependent variable to identify specific effects. Significant main effects were followed up with Tukey's HSD post-hoc tests for pairwise comparisons. The significance level was set at \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.05. All data are reported as \u003cem\u003eM ± SE\u003c/em\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResult\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eGiven that the dependent variables of interest—retaliation, understanding, and communication—were theoretically related, a one-way multivariate analysis of variance (MANOVA) was conducted to test for an overall effect of the AI's response style on the linear combination of these variables.\u003c/p\u003e\n\u003cp\u003eThe results of the MANOVA are presented in Table 1. The analysis revealed a significant multivariate main effect of condition in the repeated scenario, indicating a significant difference among the groups on the combined dependent variables. However, this multivariate effect was not significant in the parallel scenario.\u003c/p\u003e\n\u003cp\u003eTo follow up on the significant multivariate effect in the repeated scenario, a series of one-way ANOVAs were conducted separately for each of the three dependent variables. These analyses revealed a significant main effect of condition on retaliation, \u003cem\u003eF\u003c/em\u003e(2, 76) = 3.76, \u003cem\u003ep\u003c/em\u003e = 0.028, ω² = 0.07; understanding, \u003cem\u003eF\u003c/em\u003e(2, 76) = 7.17, \u003cem\u003ep\u003c/em\u003e = 0.001, ω² = 0.14; and communication, \u003cem\u003eF\u003c/em\u003e(2, 76) = 7.31, \u003cem\u003ep\u003c/em\u003e = 0.001, ω² = 0.14.\u003c/p\u003e\n\u003cp\u003ePairwise comparisons using Tukey's HSD post-hoc tests clarified the nature of these effects (see Figures 2A-C). For the retaliation variable, the positive-response (PR) group reported significantly lower intentions than the control (CON) group, while no other pairwise comparisons were significant (Fig.2A). A different pattern emerged for understanding, where both the PR and negative-response (NR) groups demonstrated significantly higher intentions than the CON group, with no significant difference found between the two AI conditions (Fig.2B). Finally, regarding the intention to communicate, the PR group scored significantly higher than both the NR and CON groups, between which there was no significant difference (Fig. 2C). The distributions of individual data points for each group are presented in Figures 2D-F.\u003c/p\u003e\n\u003ch3\u003eDiscussion1\u003c/h3\u003e\n\u003cp\u003eThe results of Study 1 demonstrate that engaging in role-playing with an AI in a simulated conflict scenario significantly enhanced participants' intentions to understand and communicate, while reducing their retaliatory tendencies. This finding supports our first hypothesis (H1). However, this positive effect did not generalize to the novel parallel context. This suggests that AI-enabled role-playing (AI-RP) may function as a context-dependent intervention, whose effects diminish once the immediate intervention context is removed.\u003c/p\u003e\n\u003cp\u003eThis phenomenon can be interpreted through the lens of Vygotsky's theory of cognitive scaffolding(Vygotsky \u0026amp; Cole, 1978). Interpersonal conflicts are inherently situations of high cognitive load(Sweller, 1988). Within the structured and simulated nature of the role-playing task, participants' cognitive load may be temporarily alleviated, enabling them to allocate more cognitive resources towards understanding their partner's perspective. In this context, the AI acts as a form of external support, providing a \"cognitive scaffold\" that temporarily reduces the burden of managing the high-load conflict situation. However, once this scaffold is removed and participants are confronted with a new parallel context, their cognitive load likely increases, causing them to revert to their habitual, more automatic modes of conflict resolution.\u003c/p\u003e\n\u003cp\u003eA deeper analysis of the different AI response styles reveals a nuanced pattern. While the positive-response (PR) group exhibited significantly higher communication intentions than the negative-response (NR) group, the two groups did not differ significantly on the dimensions of understanding and retaliation. Notably, the NR group still demonstrated significantly better outcomes on these two dimensions compared to the no-interaction control group. These results point toward two distinct mechanisms underlying the AI-RP intervention.\u003c/p\u003e\n\u003cp\u003eThe first mechanism is the activation of perspective-taking. Regardless of the AI's response style, the role-playing exercise itself provides a continuous, external interaction partner, which appears to compel participants to engage in perspective-taking(Galinsky \u0026amp; Moskowitz, 2000). The necessity of responding to the AI partner forces individuals to move beyond their own preconceived biases and consider the other's point of view, a process known to reduce animosity and enhance understanding(Galinsky et al., 2005). Thus, a fundamental function of AI-RP is to provide a safe interactive space that prompts cognitive reappraisal, thereby reducing retaliatory intentions and increasing understanding. The second mechanism is the provision of emotional support, where the AI's positive empathy plays an additional, distinct role. Initiating communication is often perceived as having high interpersonal risk. By experiencing understanding and acceptance from the positive AI, participants' sense of psychological safety is enhanced, which can significantly lower the perceived risks associated with communication(Fredrickson, 2001). This, in turn, motivates them to engage in more constructive communication behaviors.\u003c/p\u003e\n\u003cp\u003eIn summary, Study 1 provides a preliminary validation of the feasibility of AI-enabled role-playing (AI-RP) for improving conflict resolution strategies, highlighting its distinct roles in activating perspective-taking and providing emotional support. However, these findings raise a deeper theoretical question: Do the \"algorithmic empathy\" of an AI and the \"embodied empathy\" of a human differ fundamentally in nature, and does this potential difference impact the intervention's effectiveness? While some scholars argue that an AI cannot achieve genuine human empathy due to its lack of authentic emotional experience(Harel, 2004), its \"Rational Empathy\" may, conversely, offer unique advantages owing to its impartiality and consistency(Adam et al., 2021).\u003c/p\u003e\n\u003cp\u003eTherefore, Study 2 was designed to directly compare role-playing with an AI versus a human counselor, aiming to investigate which form of empathy is more conducive to promoting constructive conflict resolution. Furthermore, through the analysis of dialogue texts, we will delve into the distinct psychological mechanisms underlying these two intervention modalities. This will provide empirical evidence for understanding the differential utility of \"Rational Empathy\" versus \"Emotional Empathy\" in practical application.\u003c/p\u003e"},{"header":"Study 2","content":"\n\u003ch3\u003eStudy 2a\u003c/h3\u003e\n\u003cp\u003e\u003cstrong\u003eMethod\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eParticipants\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAn a priori power analysis was conducted using G*Power (Version 3.1) to determine the required sample size. The analysis was set for a MANOVA: Global effects, with parameters of \u0026alpha; = 0.05, power (1-\u0026beta;) = 0.80, and a medium effect size (f\u0026sup2;(V) = 0.08), which indicated a minimum required sample of 90 participants. We initially recruited 100 university students. Seventeen participants were excluded for failing to complete the full experimental procedure, resulting in a final sample of 83 individuals. Participants were randomly assigned to one of three groups: a human counselor group (n = 29, 8 males, 16 females; \u003cem\u003eM\u003c/em\u003e = 20.24, SD = 1.66), an AI group (n = 26, 7 males, 19 females; \u003cem\u003eM\u003c/em\u003e = 19.85, SD = 1.22), or a no-interaction control group (CON; n = 28, 7 males, 21 females; \u003cem\u003eM\u0026nbsp;\u003c/em\u003e= 20.46, SD = 1.32). All experimental and ethical procedures were consistent with those of Study 1.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eDesign\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA single-factor, between-subjects design was employed. The independent variable was the type of role-playing partner, with three levels: a human counsellor(Counsellor), an artificial intelligence (AI), and a no-interaction control(Control). The dependent variables were participants\u0026apos; intentions of retaliation, understanding, and communication. This study aimed to directly compare the effects of role-playing with a human counselor versus an AI on participants\u0026apos; conflict resolution intentions under a positive-response condition.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eMaterials and Measures\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eConflict Scenario\u003c/em\u003e. The conflict scenario used in this study involved an online multiplayer game: the protagonist publicly pointed out a teammate\u0026apos;s (F\u0026apos;s) tactical error that led to the team\u0026apos;s loss. In response, Teammate F publicly posted the protagonist\u0026apos;s previous term paper plagiarism report in a 200-person university major group chat, questioning their academic integrity.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eDependent Measures\u003c/em\u003e. The measures for retaliation, understanding, and communication intentions were identical to those used in Study 1, employing a 5-point Likert scale.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eCommunication and AI Tools\u003c/em\u003e. As in Study 1, interactions were conducted via QQ using the \u0026quot;Wizard of Oz\u0026quot; technique. The \u0026quot;Doubao\u0026quot; model was again used as the AI tool.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eHuman Counsellor.\u003c/em\u003e Interactions in the counselor group were conducted by a senior licensed psychotherapist with over 10 years of professional training and 7 years of clinical experience, accumulating nearly 3,000 hours of client sessions and approximately 500 hours of supervised practice.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eProcedure\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe overall experimental procedure was similar to that of Study 1, with three key differences:(1) To gather richer dialogue data for subsequent analysis (Study 2b), the role-playing interaction was extended to six turns per participant (12 messages in total); (2) Both the AI and the human counselor were instructed to provide positive, empathic responses, following identical intervention guidelines to ensure stylistic consistency; (3) Unlike Study 1, this study did not include a parallel context, aiming to focus more directly on the immediate effects of the two intervention modalities in a specific scenario. Consequently, the dependent variables were measured only in relation to the role-playing context itself. The procedure for the control (CON) group was identical to that in Study 1. All participants were fully debriefed at the conclusion of the experiment.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eData Analysis\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe quantitative data analysis approach for Study 2a (MANOVA followed by ANOVAs) was identical to that of Study 1. In addition, the dialogue texts from the two experimental groups (AI and counsellor) were coded for conflict resolution outcomes, categorized into three types: (1) Conflict Resolved (e.g., participant expressed positive emotions and willingness to reconcile); (2) Conflict Partially Resolved (e.g., participant still expressed some negativity but was hopeful about the future relationship); and (3) Conflict Unresolved (e.g., participant explicitly refused reconciliation). To ensure objectivity, two independent raters, blind to the experimental conditions, coded the outcomes. Inter-rater reliability was assessed using Cohen\u0026apos;s Kappa, which indicated a substantial level of agreement (\u0026kappa; = 0.727, \u003cem\u003ep\u0026nbsp;\u003c/em\u003e\u0026lt; 0.001, 95%\u003cem\u003e\u0026nbsp;CI\u003c/em\u003e [0.577, 0.878]).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResult\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe results of the one-way MANOVA are presented in Table 2. The analysis revealed a significant multivariate main effect of the role-playing partner on the combined dependent variables of retaliation, understanding, and communication. Subsequent one-way ANOVAs were conducted for each dependent variable. These analyses revealed a significant main effect of condition on retaliation, \u003cem\u003eF\u003c/em\u003e(2, 80) = 7.27, \u003cem\u003ep\u003c/em\u003e = 0.001, \u0026omega;\u0026sup2; = 0.13; understanding, \u003cem\u003eF\u003c/em\u003e(2, 80) = 8.25, \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.001, \u0026omega;\u0026sup2; = 0.15; and communication, \u003cem\u003eF\u003c/em\u003e(2, 80) = 8.24, \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.001, \u0026omega;\u0026sup2; = 0.15.\u003c/p\u003e\n\u003cp\u003ePairwise comparisons using Tukey\u0026apos;s HSD post-hoc tests clarified these effects (see Figures 4A-C). For retaliation, both the AI and Counsellor groups reported significantly lower intentions than the Control group, with no significant difference between the AI and Counsellor groups themselves (Fig.4A). A similar pattern was observed for understanding, where the AI and Counsellor groups both reported significantly higher intentions than the Control group, and again did not differ from each other (Fig.4B). For communication, however, the AI group scored significantly higher than both the Counsellor and Control groups, between which there was no significant difference (Fig.4C).\u003c/p\u003e\n\u003cp\u003eIn addition to the scaled measures, we analyzed the categorical conflict resolution outcomes based on the coded dialogue texts (see Figure 4D). The results were broadly consistent with the post-hoc findings. A slightly higher percentage of participants in the AI group (64.52%) reached a \u0026quot;Conflict Resolved\u0026quot; outcome compared to the Counsellor group (60.71%). The proportions of participants in the \u0026quot;Partially Resolved\u0026quot; and \u0026quot;Unresolved\u0026quot; categories were comparable across the two groups.\u003c/p\u003e\n\u003ch3\u003eStudy 2b\u003c/h3\u003e\n\u003cp\u003e\u003cstrong\u003eMethod\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eParticipants and Corpus\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe present analysis was conducted on the dialogue text data generated by the 55 participants in the AI (n = 26) and Counsellor (n = 29) groups from Study 2a. The complete dialogue corpus consisted of 660 conversational turns (55 participants × 6 turns × 2 interlocutors). For comparative analysis, the corpus was divided into two sub-corpora: participant-generated text and partner-generated text (from either the AI or the human counsellor).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eData Analysis Strategy\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo investigate differences in the linguistic style and content between the AI and human counsellor interactions, we employed two computational methods: linguistic feature analysis and topic modeling. All analyses were performed in a Python 3.10 environment using Jupyter Notebooks.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eText Preprocessing\u003c/em\u003e：The text preprocessing pipeline was adapted for the specific requirements of each analytical method. For the linguistic feature analysis (LIWC), the text underwent several steps: removal of non-textual content (e.g., emojis), punctuation removal, Chinese word segmentation using the jieba library, and stop-word removal. For the topic modeling (BERTopic), only non-textual content was removed to preserve the complete semantic structure of the sentences.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eLinguistic Feature Analysis\u003c/em\u003e：We employed the Linguistic Inquiry and Word Count (LIWC) framework to quantify psycholinguistic features in the text. The analysis was based on the official LIWC2015 Simplified Chinese dictionary (Pennebaker et al., 2015), which includes validated categories for emotional, cognitive, and social processes. A custom Python script was used to calculate the proportion of words belonging to each category relative to the total word count for each text unit (i.e., participant response or partner response), allowing for standardized comparisons between groups.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eTopic Modeling\u003c/em\u003e：To uncover the latent thematic structure of the dialogues, we utilized the BERTopic (v0.16.0) model (Grootendorst, 2022). The analysis pipeline consisted of three steps: (1) Text Vectorization: Preprocessed texts were converted into high-dimensional vectors using the SentenceTransformer model ('paraphrase-multilingual-MiniLM-L12-v2'). (2) Topic Generation: The resulting text embeddings were fed into the BERTopic model, which performed topic modeling via clustering and the c-TF-IDF algorithm. Key parameters were set to language='chinese (simplified)', nr_topics='auto', and min_topic_size=10 to automatically determine the optimal number of topics. (3) Topic Interpretation and Naming: To ensure accurate and readable interpretations of the generated topics, we adopted a human-AI collaborative approach (cf. Alharbi et al., 2024). The keyword lists and representative sentences for each topic generated by BERTopic were provided as input to a large language model, ChatGPT (based on the GPT-5 architecture), to generate concise and accurate topic names and descriptions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eStatistical Testing\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo compare language use between the AI and Counsellor groups, we conducted Chi-square tests of independence, followed by standardized residual analysis. These tests were used to determine whether there were significant differences in the relative frequencies of LIWC categories and in the distribution of documents across the identified BERTopic themes. The statistical analysis relied on the pandas (2.3.1), numpy (2.3.2), and SciPy (1.14.1) Python libraries.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResult\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLinguistic Analysis of Role-Playing Interaction Texts\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo further explore the psychological mechanisms underlying the observed differences in communication intention between the AI and Counsellor groups, we analyzed the dialogue texts generated during the role-playing interactions. Specifically, we examined whether the linguistic styles of the AI and the human counsellor differed, and whether participants, in turn, used language differently when interacting with an AI versus a human partner.\u003c/p\u003e\n\u003cp\u003eOur analysis, conducted using a custom Python script, revealed distinct differences in the Linguistic Inquiry and Word Count (LIWC) category usage between the AI and the counsellor. As illustrated in Figure 5A, which displays the proportions of several key LIWC categories, the largest discrepancies were observed in the categories of Affective Processes, Function Words, Informal Language, and Social Processes. Word usage in other categories was comparatively similar. Subsequent Chi-square tests confirmed that the differences between the AI and the counsellor in these four categories were statistically significant (see Table 3 for details). We then compared the linguistic features of the participants' own texts from the two experimental groups. The results indicated that despite interacting with different partners (AI vs. counsellor), participants themselves showed no significant differences in their use of LIWC word categories (see Figure 5B and Table 4).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTopic Modeling of Role-Playing Interaction Texts\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe further employed BERTopic to conduct topic modeling on the interaction texts, aiming to explore potential differences in the thematic content of the dialogues between the experimental groups.\u003c/p\u003e\n\u003cp\u003eThe topic modeling of the partner-generated texts (from the AI and the counsellor) yielded three distinct topics. Figure 6 (left) displays the top five keywords for each topic. Based on these keywords and representative texts, and using the previously described GPT-5 assisted method, the three topics were named as follows (see also Supplementary Table 1): Topic 0: Sincere Apology and Self-Reflection (characterized by keywords related to apology and introspection); Topic 1: Intensifying Regret and Alleviating Tension (marked by expressions of remorse and efforts to de-escalate); and Topic 2: Emotional Affirmation and Relationship Repair (encompassing keywords related to validating feelings and hopes for reconciliation). We then analyzed the distribution of AI- versus counsellor-generated texts across these three topics (see Figure 6A). A Chi-square test revealed a significant difference in this distribution (see Figure 6C for standardized residuals). Specifically, the AI's responses were significantly more likely to fall under Topic 0 (Sincere Apology and Self-Reflection) compared to the counsellor's. Conversely, the counsellor's responses were significantly more concentrated in Topic 2 (Emotional Affirmation and Relationship Repair). The two partners did not differ significantly in their use of Topic 1.\u003c/p\u003e\n\u003cp\u003eThe analysis of participant-generated texts yielded four distinct topics, with the top five keywords for each shown in Figure 6 (right). Following the same interpretation procedure (see also Supplementary Table 2), these four topics were named: Topic 0: Emotional Reaction and Privacy Protection (focusing on expressions of feeling and defending personal boundaries); Topic 1: Brief Responses and Emotional Buffering (characterized by short, tension-reducing replies); Topic 2: Accepting Apology and Setting Boundaries (containing keywords related to accepting apologies while establishing future expectations); and Topic 3: Understanding, Acceptance, and Relationship Repair (focusing on reconciliation and moving forward). Finally, we compared the distribution of topics used by participants who interacted with the AI versus those who interacted with the counsellor (see Figure 6B). A Chi-square test revealed no significant difference in the topic distribution between the two groups (see Figure 6D), indicating that participants used similar thematic content regardless of their interaction partner.\u003c/p\u003e\n\u003cp\u003eWe employed topic similarity analysis and visualizations of the topic spatial distribution to visually inspect the quality of the topic clusters. For the partner-generated texts, although the topic similarity analysis indicated some overlap between Topic 0 and Topic 1 (see Supplementary Figure 1, left), the visualization of the topic space showed that the clusters were well-differentiated (see Supplementary Figure 2, left). Furthermore, a qualitative review of the semantic content of Topic 0 and Topic 1 confirmed that they represented distinct concepts. Therefore, we retained both topics rather than merging them. Similarly, for the four topics generated from the participant texts, the analyses showed both a degree of thematic similarity and clear differentiation between the clusters (see Supplementary Figure 1, right, and Supplementary Figure 2, right).\u003c/p\u003e\n\u003ch3\u003eDiscussion2\u003c/h3\u003e\n\u003cp\u003eThe results of Study 2a replicated a key finding from Study 1: role-playing, whether with an AI or a human counsellor, led to significant improvements in participants' retaliatory intentions and levels of understanding compared to the control group. This result provides robust support for the generalizability of structured role-playing as an intervention paradigm. It suggests that any medium capable of providing external interaction and feedback—thereby disrupting an individual's egocentric perspective and prompting perspective-taking—will be more effective than mere internal rumination. This further confirms the immense potential of AI to function as a \"cognitive scaffold\" for practicing conflict resolution skills. Strikingly, however, when directly comparing the intervention effects of the AI against the human expert, we found that while the two were equally effective at reducing retaliation and increasing understanding, the AI was significantly superior to the human expert in the crucial dimension of fostering communication intention.\u003c/p\u003e\n\u003cp\u003eThis finding not only supports recent research by Ayers et al. (Ayers et al., 2023)but also offers a counter-intuitive perspective: in certain conflict situations, the \"algorithmic empathy\" of an AI, which lacks embodied emotional experience, may possess unique advantages. This superiority does not stem from an AI's ability to \"feel\" in the same way humans do. Rather, it likely originates from its vast underlying data foundation, which enables it to provide efficient, consistent, and unbiased feedback. This allows the AI to offer precise, problem-focused insights and solutions that may be difficult for even a human expert to consistently match(Olawade et al., 2024).\u003c/p\u003e\n\u003cp\u003eTo delve into the underlying mechanisms of the AI's superior performance, we conducted an in-depth analysis of the dialogue texts using LIWC and BERTopic. The results revealed the central role of affective responses in both role-playing conditions; both the AI and the human counsellor used a high proportion of emotional words, confirming that affective feedback is an indispensable component of conflict mediation. However, a key difference emerged: the AI's language contained a significantly higher proportion of function words, whereas the human counsellor's language was richer in affective words.\u003c/p\u003e\n\u003cp\u003eThis aligns with the theoretical distinction between the two primary components of empathy: affective and cognitive(Deutsch \u0026amp; Madle, 1975). Affective empathy is typically defined as a vicarious, instinctual response to another's emotional state, whereas cognitive empathy involves intentionally taking on another's perspective to infer their thoughts and feelings(Shamay-Tsoory, 2011). Although the AI lacks embodied feelings, it technically recognizes and provides affective feedback to the user's emotional expressions. Crucially, compared to the human counsellor, it places a greater focus on functional responses, offering objective, problem-solving advice and suggestions. We define this specific form of empathy as \"Rational Empathy.\" This focus on providing tangible advice may have sent a potent signal to participants that \"communication is a viable pathway to resolving the problem,\" thereby increasing their motivation to engage in it.\u003c/p\u003e\n\u003cp\u003eFurthermore, the analysis revealed no significant differences in the overall linguistic profiles of the participants themselves across the two groups. This null finding helps to validate the effectiveness of our blind experimental manipulation, suggesting that confounding factors such as \"algorithm aversion\" did not significantly interfere with the results(Jussupow et al., 2020). When we focus further on the differences in response styles, a clear pattern emerges: the human counsellor's dialogue was more oriented towards \"Emotional Affirmation and Relationship Repair.\" This aligns with the principles of person-centered or humanistic training, which emphasizes the client's deeper feelings and emotional experiences, prioritizing the emotional harm caused over the situational facts of the conflict itself(Rogers, 2012).\u003c/p\u003e\n\u003cp\u003eIn contrast, the AI's responses demonstrated a more functional and goal-oriented approach. Its dialogue content was more heavily focused on \"Sincere Apology and Self-Reflection\" and \"Intensifying Regret and Alleviating Tension.\" The AI's strength lies in its high efficiency, rapidly de-escalating the conflict's intensity and using non-confrontational responses to lower the participant's sense of psychological threat. The human counsellor's strength, conversely, lies in processing the participant's deeper feelings; however, due to the limited length of the conversation and the immediacy of the outcome measurement, the benefits of this deeper approach may not have been fully captured in our results.\u003c/p\u003e\n\u003cp\u003eThe findings from Study 2b provide robust theoretical support for the AI's superiority in fostering communication intention. The AI's \"Rational Empathy\", through its efficient and non-confrontational replies, rapidly de-escalated the conflict and sent a potent underlying signal to the participant: \"communication is a fast and effective way to solve this problem.\" This problem-focused and action-oriented style may be particularly effective at motivating communication within the context of an immediate conflict. In contrast, while a human counsellor can process deep-seated emotions, the impact of this approach may have a lagged effect that is not fully manifested in a brief, time-limited interaction.\u003c/p\u003e"},{"header":"General discussion","content":"\u003cp\u003eThis research aimed to explore the potential of AI-enabled role-playing (AI-RP) in managing interpersonal conflicts and to compare its effectiveness with traditional human-led role-playing. Across two experimental studies, we not only validated the feasibility of AI-RP as an effective intervention tool but also uncovered its unique advantages and the underlying mechanisms driving them.\u003c/p\u003e\n\u003cp\u003eOur findings indicate that structured role-playing is a potent method for breaking ingrained patterns of conflict. Study 1 demonstrated that AI-RP significantly improved participants' interpersonal conflict resolution strategies, specifically by enhancing understanding and communication while reducing retaliatory intentions. Study 2a further consolidated this finding, showing that role-playing with either an AI or a human expert was superior to solitary rumination. This provides strong evidence that any form of structured role-playing intervention, by offering an external and continuously interactive partner, can effectively disrupt the egocentric thinking that is habitual in conflicts. This interactive process compels a \"perspective-taking\" mechanism, and this forced cognitive reappraisal is core to prompting more constructive responses. It is crucial to note, however, that this positive effect did not generalize to a new, parallel context. This suggests that AI-RP, in its current form, may function as a context-dependent \"cognitive scaffold.\" Future applications must explore how to help users internalize these learned skills to handle real-world, high-load conflict situations.\u003c/p\u003e\n\u003cp\u003eAnother key finding is the unique advantage of AI-RP in fostering participants' willingness to communicate. Study 2a revealed that the AI was even more effective than a human expert at enhancing communication intentions. Our text analysis uncovered the mechanism behind this counter-intuitive finding. The AI's superiority stems not from superior emotionality but from its unique combination of \"Rational Empathy\" and a problem-focused approach. Although the AI also used a high proportion of affective words, its dialogue patterns were more efficiency-oriented. By rapidly acknowledging fault and offering solutions, it sent an implicit signal that \"communication can solve this problem.\" This approach effectively lowered the participants' sense of psychological threat, making them more willing to initiate communication. This stands in stark contrast to the human expert's style, which, being more focused on emotional processing, may have a less immediate impact.\u003c/p\u003e\n\u003cp\u003eCrucially, the findings of this research are not intended to suggest that AI will fully replace human counselors but rather to highlight the value of their complementary coexistence. The topic analysis in Study 2b further substantiates this point. The human counselor's dialogue was more oriented towards \"Emotional Affirmation and Relationship Repair,\" which aligns with the deep emotional processing emphasized in humanistic therapies, but its full effects may require a longer timeframe to manifest(Howard et al., 1986). In contrast, the AI's advantage lies in its efficient, non-confrontational handling of immediate conflicts, rapidly de-escalating their intensity and creating favorable conditions for problem-solving. In future practice, AI could serve as an initial intervention tool to help individuals quickly de-escalate a conflict and build the willingness to communicate. Human experts could then build upon this foundation to provide deeper emotional support and relational repair. This human-AI collaborative model, rather than a simple substitution of humans by AI, could provide more comprehensive, efficient, and scalable solutions for resolving interpersonal conflict(Devillers, 2021).\u003c/p\u003e\n\u003cp\u003eThis research provides a solid empirical foundation for AI-assisted psychological interventions and interpersonal skills training. We have demonstrated that in dynamic role-playing, an AI can act as a novel \"cognitive scaffold\" to activate users' perspective-taking in conflicts. Furthermore, we have re-examined the concept of AI \"empathy\" by proposing the construct of \"Rational Empathy,\" offering a new direction for the design of AI in the field of Human-Computer Interaction (HCI). AI-RP can be developed into various applications, such as a simulation training tool for workplace mediation, a practice platform for family communication, or an educational tool for developing conflict resolution skills in adolescents. Its efficiency, accessibility, and scalability make large-scale, personalized psychological support a tangible possibility.\u003c/p\u003e\n\u003cp\u003eNevertheless, this study has several limitations that should be acknowledged. The lack of transferability of the observed AI-RP effects suggests that future research should explore how longer-term training might enhance skill internalization. Second, our assessment was based on immediate behavioral intentions and did not capture the long-term effects of the intervention. Additionally, the individual characteristics of the single human expert may limit the generalizability of our conclusions. Finally, cross-cultural factors, such as the emphasis on harmony in Chinese culture(Leung et al., 2011), and their potential impact on the results warrant further investigation. Future research could also explore the development of multimodal AI-RP systems that incorporate voice and facial expression recognition to enhance realism(Li \u0026amp; Deng, 2020). Moreover, investigating how to tailor the AI's intervention strategies based on users' personality traits presents an exciting avenue for future inquiry. \u0026nbsp;\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eOur research is conducted in strict accordance with the ethical principles outlined in the Declaration of Helsinki, ensuring the utmost protection of the safety, privacy, and rights of all participants. The studies involving humans were approved by the Ethics Committee of Anhui University of Science and Technology. The studies were conducted in accordance with the local legislation and institutional requirements. The entire process adhered to the principle of voluntary participation. Participants were free to choose to withdraw from the experiment. After the study was completed, the investigator clearly explained the purpose of the study to all participants.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll data generated or analyzed during this study are included in this published article.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the University-Level Quality Engineering Project of Anhui Science and Technology University (Project No. Xj2024200).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors' contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTTY: Conceptualization, Formal analysis, Investigation, Writing – original draft, Funding acquisition. HLM: Conceptualization, Methodology. \u0026nbsp;HYW: Supervision. HDS: Conceptualization, Project administration, Methodology, Supervision, Writing – original draft, Writing – review \u0026amp; editing.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e We sincerely acknowledge all participants involved in this research and those who assisted in the recruitment process. This study was supported by the University-Level Quality Engineering Project of Anhui University of Science and Technology (Project No. Xj2024200).\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAchiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., \u0026amp; Anadkat, S. (2023). Gpt-4 technical report. \u003cem\u003earXiv preprint arXiv:2303.08774\u003c/em\u003e. \u003c/li\u003e\n\u003cli\u003eAdam, M., Wessel, M., \u0026amp; Benlian, A. (2021). Ai-based chatbots in customer service and their effects on user compliance. \u003cem\u003eElectronic Markets\u003c/em\u003e,\u003cem\u003e 31\u003c/em\u003e(2), 427-445. https://doi.org/10.1007/s12525-020-00414-7 \u003c/li\u003e\n\u003cli\u003eAdamopoulou, E., \u0026amp; Moussiades, L. (2020). An overview of chatbot technology. IFIP international conference on artificial intelligence applications and innovations, \u003c/li\u003e\n\u003cli\u003eAydoğan, R., Baarslag, T., \u0026amp; Gerding, E. (2021). Artificial intelligence techniques for conflict resolution. \u003cem\u003eGroup Decision and Negotiation\u003c/em\u003e,\u003cem\u003e 30\u003c/em\u003e(4), 879-883. https://doi.org/10.1007/s10726-021-09738-x \u003c/li\u003e\n\u003cli\u003eAyers, J. W., Poliak, A., Dredze, M., Leas, E. C., Zhu, Z., Kelley, J. B., Faix, D. J., Goodman, A. M., Longhurst, C. A., Hogarth, M., \u0026amp; Smith, D. M. (2023). Comparing physician and artificial intelligence chatbot responses to patient questions posted to a public social media forum. \u003cem\u003eJAMA Internal Medicine\u003c/em\u003e,\u003cem\u003e 183\u003c/em\u003e(6), 589-596. https://doi.org/10.1001/jamainternmed.2023.1838 \u003c/li\u003e\n\u003cli\u003eChaturvedi, A. (2024). Exploring empathy in artificial intelligence: Synthesis and paths for future research. \u003cem\u003eInformation Discovery and Delivery\u003c/em\u003e,\u003cem\u003e 53\u003c/em\u003e(3), 406-424. https://doi.org/10.1108/IDD-03-2024-0048 \u003c/li\u003e\n\u003cli\u003eDavis, M. H. (1983). Measuring individual differences in empathy: Evidence for a multidimensional approach. \u003cem\u003eJournal of Personality and Social Psychology\u003c/em\u003e,\u003cem\u003e 44\u003c/em\u003e(1), 113. https://doi.org/10.1037//0022-3514.44.1.113 \u003c/li\u003e\n\u003cli\u003eDeutsch, F., \u0026amp; Madle, R. A. (1975). Empathy: Historic and current conceptualizations, measurement, and a cognitive theoretical perspective. \u003cem\u003eHuman development\u003c/em\u003e,\u003cem\u003e 18\u003c/em\u003e(4), 267-287. https://doi.org/10.1159/000271488 \u003c/li\u003e\n\u003cli\u003eDevillers, L. (2021). Human\u0026ndash;robot interactions and affective computing: The ethical implications. In \u003cem\u003eRobotics, ai, and humanity: Science, ethics, and policy\u003c/em\u003e (pp. 205-211). Springer International Publishing Cham. \u003c/li\u003e\n\u003cli\u003eFeldkamp, T., Langer, M., Wies, L., \u0026amp; K\u0026ouml;nig, C. J. (2023). Justice, trust, and moral judgements when personnel selection is supported by algorithms. \u003cem\u003eEuropean Journal of Work and Organizational Psychology\u003c/em\u003e,\u003cem\u003e 33\u003c/em\u003e(2), 130-145. https://doi.org/10.1080/1359432x.2023.2169140 \u003c/li\u003e\n\u003cli\u003eFigley, C. R. (2002). Compassion fatigue: Psychotherapists\u0026apos; chronic lack of self care. \u003cem\u003eJournal of clinical psychology\u003c/em\u003e,\u003cem\u003e 58\u003c/em\u003e(11), 1433-1441. https://doi.org/10.1002/jclp.10090 \u003c/li\u003e\n\u003cli\u003eFredrickson, B. L. (2001). The role of positive emotions in positive psychology: The broaden-and-build theory of positive emotions. \u003cem\u003eAmerican Psychologist\u003c/em\u003e,\u003cem\u003e 56\u003c/em\u003e(3), 218-226. https://doi.org/10.1037/0003-066X.56.3.218 \u003c/li\u003e\n\u003cli\u003eGalinsky, A. D., Ku, G., \u0026amp; Wang, C. S. (2005). Perspective-taking and self-other overlap: Fostering social bonds and facilitating social coordination. \u003cem\u003eGroup processes \u0026amp; intergroup relations\u003c/em\u003e,\u003cem\u003e 8\u003c/em\u003e(2), 109-124. https://doi.org/10.1177/1368430205051060 \u003c/li\u003e\n\u003cli\u003eGalinsky, A. D., \u0026amp; Moskowitz, G. B. (2000). Perspective-taking: Decreasing stereotype expression, stereotype accessibility, and in-group favoritism. \u003cem\u003eJournal of Personality and Social Psychology\u003c/em\u003e,\u003cem\u003e 78\u003c/em\u003e(4), 708-724. https://doi.org/10.1037/0022-3514.78.4.708 \u003c/li\u003e\n\u003cli\u003eHarel, D. (2004). \u003cem\u003eComputers ltd: What they really can\u0026apos;t do\u003c/em\u003e. OUP Oxford. \u003c/li\u003e\n\u003cli\u003eHoward, K. I., Kopta, S. M., Krause, M. S., \u0026amp; Orlinsky, D. E. (1986). The dose\u0026ndash;effect relationship in psychotherapy. \u003cem\u003eAmerican Psychologist\u003c/em\u003e,\u003cem\u003e 41\u003c/em\u003e(2), 159. https://doi.org/10.1037/0003-066X.41.2.159 \u003c/li\u003e\n\u003cli\u003eHsu, A., \u0026amp; Chaudhary, D. (2023). Ai4pcr: Artificial intelligence for practicing conflict resolution. \u003cem\u003eComputers in Human Behavior: Artificial Humans\u003c/em\u003e,\u003cem\u003e 1\u003c/em\u003e(1), 100002. https://doi.org/https://doi.org/10.1016/j.chbah.2023.100002 \u003c/li\u003e\n\u003cli\u003eJia, F. (2014). \u003cem\u003eThe impact of heterogeneous high self-esteem on college students\u0026apos; interpersonal conflict resolution style under situational forgiveness\u003c/em\u003e NanJing Normal University]. \u003c/li\u003e\n\u003cli\u003eJussupow, E., Benbasat, I., \u0026amp; Heinzl, A. (2020). Why are we averse towards algorithms? A comprehensive literature review on algorithm aversion. \u003c/li\u003e\n\u003cli\u003eKalla, D., Smith, N., Samaah, F., \u0026amp; Kuraku, S. (2023). Study and analysis of chat gpt and its impact on different fields of study. \u003cem\u003eInternational journal of innovative science and research technology\u003c/em\u003e,\u003cem\u003e 8\u003c/em\u003e(3). \u003c/li\u003e\n\u003cli\u003eKim, T., \u0026amp; Peng, W. (2024). Do we want ai judges? The acceptance of ai judges\u0026rsquo; judicial decision-making on moral foundations. \u003cem\u003eAI \u0026amp; SOCIETY\u003c/em\u003e,\u003cem\u003e 40\u003c/em\u003e(5), 3683-3696. https://doi.org/10.1007/s00146-024-02121-9 \u003c/li\u003e\n\u003cli\u003eKlimecki, O. M. (2019). The role of empathy and compassion in conflict resolution. \u003cem\u003eEmotion Review\u003c/em\u003e,\u003cem\u003e 11\u003c/em\u003e(4), 310-325. https://doi.org/10.1177/1754073919838609 \u003c/li\u003e\n\u003cli\u003eLeung, K., Brew, F. P., Zhang, Z.-X., \u0026amp; Zhang, Y. (2011). Harmony and conflict: A cross-cultural investigation in china and australia. \u003cem\u003eJournal of Cross-Cultural Psychology\u003c/em\u003e,\u003cem\u003e 42\u003c/em\u003e(5), 795-816. https://doi.org/10.1177/0022022110363474 \u003c/li\u003e\n\u003cli\u003eLevy, K. N., \u0026amp; Scala, J. W. (2012). Transference, transference interpretations, and transference-focused psychotherapies. \u003cem\u003ePsychotherapy\u003c/em\u003e,\u003cem\u003e 49\u003c/em\u003e(3), 391-403. https://doi.org/10.1037/a0029371 \u003c/li\u003e\n\u003cli\u003eLi, S., \u0026amp; Deng, W. (2020). Deep facial expression recognition: A survey. \u003cem\u003eIEEE transactions on affective computing\u003c/em\u003e,\u003cem\u003e 13\u003c/em\u003e(3), 1195-1215. https://doi.org/10.1109/TAFFC.2020.2981446 \u003c/li\u003e\n\u003cli\u003eLutgen-Sandvik, P., \u0026amp; Sypher, B. D. (2009). Destructive organizational communication. \u003cem\u003eProcesses, Consequences and Constructive Ways of Organizing\u003c/em\u003e. \u003c/li\u003e\n\u003cli\u003eMead, G. H. (2015). \u003cem\u003eMind, self \u0026amp; society\u003c/em\u003e. University of Chicago Press. \u003c/li\u003e\n\u003cli\u003eOlawade, D. B., Wada, O. Z., Odetayo, A., David-Olawade, A. C., Asaolu, F., \u0026amp; Eberhardt, J. (2024). Enhancing mental health with artificial intelligence: Current trends and future prospects. \u003cem\u003eJournal of Medicine, Surgery, and Public Health\u003c/em\u003e,\u003cem\u003e 3\u003c/em\u003e, 100099. https://doi.org/10.1016/j.glmedi.2024.100099 \u003c/li\u003e\n\u003cli\u003ePagani, A. F., Donato, S., Parise, M., Bertoni, A., Iafrate, R., \u0026amp; Schoebi, D. (2019). Explicit stress communication facilitates perceived responsiveness in dyadic coping. \u003cem\u003eFrontiers in psychology\u003c/em\u003e,\u003cem\u003e 10\u003c/em\u003e, 401. https://doi.org/10.3389/fpsyg.2019.00401 \u003c/li\u003e\n\u003cli\u003eRogers, C. (2012). \u003cem\u003eClient centered therapy (new ed)\u003c/em\u003e. Hachette UK. \u003c/li\u003e\n\u003cli\u003eRosler, N., Cohen-Chen, S., \u0026amp; Halperin, E. (2015). The distinctive effects of empathy and hope in intractable conflicts. \u003cem\u003eSage Journals\u003c/em\u003e. https://doi.org/10.1177/0022002715569772 \u003c/li\u003e\n\u003cli\u003eShamay-Tsoory, S. G. (2011). The neural bases for empathy. \u003cem\u003eThe Neuroscientist\u003c/em\u003e,\u003cem\u003e 17\u003c/em\u003e(1), 18-24. https://doi.org/10.1177/1073858410379268 \u003c/li\u003e\n\u003cli\u003eSweller, J. (1988). Cognitive load during problem solving: Effects on learning. \u003cem\u003eCognitive Science\u003c/em\u003e,\u003cem\u003e 12\u003c/em\u003e(2), 257-285. https://doi.org/10.1016/0364-0213(88)90023-7 \u003c/li\u003e\n\u003cli\u003eTaguchi, N. (2023). Using immersive virtual reality for the assessment of intercultural conflict mediation. \u003cem\u003eLanguage Learning \u0026amp; Technology\u003c/em\u003e,\u003cem\u003e 27\u003c/em\u003e(3), 108-128. https://doi.org/10.64152/10125/73534 \u003c/li\u003e\n\u003cli\u003eThoits, P. A. (2011). Mechanisms linking social ties and support to physical and mental health. \u003cem\u003eJournal of health and social behavior\u003c/em\u003e,\u003cem\u003e 52\u003c/em\u003e(2), 145-161. https://doi.org/10.1177/0022146510395592 \u003c/li\u003e\n\u003cli\u003eVygotsky, L. S., \u0026amp; Cole, M. (1978). \u003cem\u003eMind in society: Development of higher psychological processes\u003c/em\u003e. Harvard University Press. \u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Anhui Science and Technology University","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, AI, Human Counsellor, Empathy, Conflict","lastPublishedDoi":"10.21203/rs.3.rs-7699927/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7699927/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eWith the rise of large language models (LLMs), LLM-driven artificial intelligence role-playing (AI-RP) has emerged as a promising tool for resolving interpersonal conflict. However, the differential effectiveness of its \"Rational Empathy\" compared to human empathy remains unclear. This study investigates this issue through two experiments. Study 1 examined the effects of different AI response styles, while Study 2 directly compared interventions by an AI and a human counsellor, incorporating linguistic analysis (LIWC and BERTopic) to investigate the underlying mechanisms. Results revealed that while AI-RP effectively improved conflict resolution outcomes, this effect was context-dependent and did not generalize. Crucially, on the key metric of communication intention, the AI was significantly superior to the human counsellor. Linguistic analysis indicated that the AI\u0026rsquo;s responses were more focused on functional, problem-solving approaches, whereas the counsellor\u0026rsquo;s focused more on affective and relational aspects. This research demonstrates that an AI can act as a \"cognitive scaffold\" in conflicts. Its unique advantage stems from an efficient, problem-oriented \"Rational Empathy\" that signals the viability of communication, offering a new perspective for future human-AI collaborative interventions.\u003c/p\u003e","manuscriptTitle":"AI's Rational Empathy Promotes Reconciliation in Conflict: Evidence from Behavioral Experiments, Linguistic Analysis, and Topic Modeling","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-26 07:00:32","doi":"10.21203/rs.3.rs-7699927/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":"20b6c4fb-a7d6-4df6-b3ca-5ac3a6ed0921","owner":[],"postedDate":"September 26th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":55239173,"name":"Psychology"}],"tags":[],"updatedAt":"2025-09-26T07:00:32+00:00","versionOfRecord":[],"versionCreatedAt":"2025-09-26 07:00:32","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7699927","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7699927","identity":"rs-7699927","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}