QualiPilot – an offline AI tool for qualitative content analysis

QualiPilot – an offline AI tool for qualitative content analysis | 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 QualiPilot – an offline AI tool for qualitative content analysis Christian Kempny, Kübra Annac, Yuece Yilmaz-Aslan, Patrick Brzoska This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7122121/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 Background The use of artificial intelligence (AI) in qualitative content analysis has increased substantially. However, most AI-assisted tools rely on proprietary cloud-based solutions, raising significant concerns regarding data protection, transparency, and reproducibility. Sensitive qualitative data such as interview transcripts require strict data sovereignty and compliance with legal and ethical standards, which existing cloud-based AI platforms cannot always guarantee. To address these challenges, this study aimed to develop and evaluate an accessible, fully offline AI tool tailored for qualitative research. Methods We developed QualiPilot, a stand-alone, privacy-compliant desktop application for AI-assisted qualitative analysis. The tool operates entirely offline, leveraging a locally deployed large language model (LLM) and requiring no installation, as it can be run directly from a USB flash drive. QualiPilot enables key steps of qualitative analysis, including initial text engagement, inductive category development, coding, and code summarization, with complete transparency of analytic parameters, system prompts, and procedural steps. Results QualiPilot performs all key analytic tasks on standard office computers without transmitting data to external servers, ensuring complete data protection and transparency. The tool produces relevant and well-structured outputs, with all steps, parameters, and prompts fully documented and exportable for reporting. The offline approach also simplifies compliance with data protection and ethical requirements. However, processing speed is slower compared to online or GPU-accelerated solutions, particularly for large datasets, and longer transcripts require manual segmentation due to model limitations. Despite these constraints, no instability or data loss is observed, and the tool proves practical for most typical qualitative projects. Conclusions QualiPilot provides a novel, privacy-compliant, and fully offline solution for AI-assisted qualitative content analysis. By enabling transparent, reproducible, and secure analytic workflows on standard hardware, the tool lowers barriers for researchers working with sensitive data. While not a replacement for manual analysis, QualiPilot offers valuable supplementary perspectives and supports rigorous, transparent qualitative research without dependence on commercial cloud services. Offline AI-assisted qualitative analysis Qualitative Research Thematic Analysis Background Interest in AI-assisted qualitative content analysis has increased in recent years, with researchers exploring a range of approaches from traditional enhancements of established qualitative methods to the development of entirely new, AI-driven analytical frameworks [1–3]. The use of LLMs in qualitative research offers significant opportunities but also entails substantial risks that span methodological, ethical, and epistemological domains. Key concerns include systematic bias introduced by LLMs, which can lead to distorted or misleading findings, particularly for sensitive topics or marginalized groups [4–6]. LLMs are prone to generating “hallucinations” (plausible but factually incorrect or fabricated content), often presented with confidence, increasing the risk of uncritical acceptance by researchers [5, 7]. Moreover, LLM-generated outputs may lack contextual depth and overlook important cultural or individual nuances, thereby affecting the credibility and richness of qualitative analysis [6, 8]. To mitigate these risks, it is essential to ensure critical human oversight, maintain transparency in LLM use and prompt design, and uphold strict ethical and data protection standards. Still, despite these shortcomings, LLMs have been shown to serve as valuable research partners by providing complementary insights that support and enrich the research process [13–15]. Additionally, LLMs can reveal patterns and connections often overlooked in manual review, particularly when researchers “converse” with their materials by asking targeted questions, probing for contextual insights, and refining theoretical connections [16, 17]. However, a persistent challenge in exploiting that potential is that most practical solutions currently rely on cloud-based infrastructures, such as MAXQDA AI Assist [9], ATLAS.ti [10] and various chatbot-based platforms like OpenAI or Meta. While these tools make powerful language models accessible for qualitative analysis, they raise substantial concerns regarding data protection, transparency, and compliance with research ethics. Sensitive qualitative data, particularly interview transcripts and personal narratives, are often subject to stringent legal and ethical requirements that mandate the researcher’s direct control over the data and limits its transfer to external servers [11–14]. Moreover, a significant limitation of existing online solutions is their lack of transparency and reproducibility. Many cloud-based AI tools do not disclose critical information such as which large language model (LLM) and version is being used, which parameters and prompts are applied, or how system prompts influence the analytic output [15, 16]. However, prompting strategies as well as the version and configuration of an LLM can significantly influence analytic outcomes [17–20]. Such details are rarely disclosed in the reporting of qualitative studies that utilize AI support [21]. This opacity complicates both the reporting and replication of research findings, raising concerns about scientific validity, methodological rigor, and ethical and legal accountability regarding data transparency. The integration of offline, locally deployed AI models into qualitative data analysis tools offers a compelling response to these limitations. Offline large language models (Local LLMs) can be run directly on researchers’ computers, enabling complete data sovereignty, maximum transparency, and reproducibility [22]. In addition, all analytic steps, model parameters, and prompt designs can be fully documented and controlled by the research team, supporting comprehensive, transparent reporting and robust reproducibility of findings [23]. Methodologically, the integration of local LLMs enables several new opportunities in qualitative analysis. From a quality assurance perspective, local AI systems enable rigorous validation processes. Given evidence showing that AI-assisted qualitative analysis can support researchers [2, 24, 25], the availability of local LLMs and structured reporting of all information makes these analyses reproducible without dependence on online solutions that do not disclose key information or on tools that modify the LLM through updates. Finally, from a legal and ethical standpoint, local AI solutions simplify compliance with data protection regulations and institutional ethics requirements. Because no data leaves the researcher’s environment, and all processing steps can be precisely documented, informed consent and ethics approvals are easier to secure and substantiate for review boards and funding agencies. With the rise of edge computing and quantized models, powerful LLMs are becoming increasingly accessible on standard hardware. A key advantage of local deployment is that the LLM remains permanently available in a fixed version because updates occur only with explicit model releases. This provides consistent access and reproducible outputs, which is not guaranteed when using cloud-based services where automatic updates can unpredictably change model behavior and affect performance. However, running large language models locally also introduces practical challenges. Users may face barriers such as insufficient administrative rights or complicated installation processes, particularly within institutional environments. In addition, limited familiarity with technical concepts such as model selection, prompt engineering, or system prompts can further hinder effective use in research settings. Designing appropriate prompts tailored to the specific LLM is therefore essential to enable meaningful and reliable application. Considering these challenges, this study aims to develop an application-friendly, local LLM solution (‘QualiPilot‘) tailored explicitly for qualitative research. The goal is to provide a tool that not only functions entirely on local hardware but also prioritizes user accessibility through an intuitive interface and the ability to run directly from a USB flash drive without installation. By focusing on comprehensive and transparent reporting of all analytic parameters and system prompts, the tool enables full reproducibility and methodological transparency. Importantly, this solution is designed to be freely available for use in academic and research contexts, thereby lowering the entry barriers for researchers who may lack access to commercial or cloud-based AI services or are hesitant to use them. Methods The QualiPilot application was developed as a stand-alone, privacy-compliant, and fully offline desktop tool for AI-assisted qualitative content analysis. In contrast to traditional CAQDAS platforms, QualiPilot operates independently as a chat-based application. Its design enables researchers to perform key steps of inductive qualitative analysis, including initial text engagement, inductive category development, coding, and code summarization, by interacting with a local LLM. The technical implementation of QualiPilot prioritizes accessibility and transparency. The tool is written in Python and uses the llama_cpp library [26] to interface with a small llama-2 model, which can be stored and executed locally with minimal RAM requirements. All computations and text processing are performed entirely on the user's computer, ensuring that no data is transmitted to external servers. The application is optimized for standard hardware, requiring a minimum of 8 GB of RAM and 4 threads, making it broadly accessible to most users. For high-performance systems, a larger and more advanced model variant will be available, though this exceeds the resource capacity of many typical PCs. A key feature of QualiPilot is its ability to apply tailored parameter settings for each analytical step, dynamically adjusting these settings to suit both the analytical task at hand and the capabilities of the chosen LLM model. For each step—initial text engagement, category development, coding, and code summarization—specific model settings are used to optimize performance and analytic quality. The parameter settings for each analytic task are shown in Table 1 . The user interface makes all model and parameter information transparent, encouraging the documentation of these settings for reproducibility. Table 1 LLM Model Parameters Used in QualiPilot for Different Analytic Tasks Analytic Task Temperature Top-k Top-p Min-p Repeat Penalty Max Answer Tokens Max Input Tokens Initial text engagement / Memoing 0.9 50 0.95 0.05 1.0 1024 4096 Inductive category development 0.7 50 0.95 0.05 1.0 2048 4096 Coding (individual segments) 0.8 50 0.95 0.05 1.1 512 4096 Code summarization 0.7 50 0.95 0.05 1.0 2048 4096 Note. Each parameter controls a specific aspect of text generation. Temperature (usually 0.1–2.0) sets how creative or random the answers are. Higher temperature values lead to more creative, lower values to more predictable responses. Top-k (1–100) means the model will only choose from the k most likely next words instead of considering all possibilities. Top-p (0–1) works a bit differently: the model considers the most likely next words until their total probability adds up to p, and then picks from this group. Min-p (0–1) filters out words that are too unlikely. Repeat Penalty (usually 1.0–2.0) helps avoid repeating the same words or phrases. Max Answer Tokens sets how long the answer can be, and Max Input Tokens limits how much text can be input at once [27]. All analytic tasks are guided by specialized system prompts tailored for qualitative research. The workflow is designed to enable users to conduct initial reading and develop a unified inductive category system, assign codes, and create code summaries, with each step’s outputs being fully documented. All analytical steps and settings are listed and can be exported for reporting and quality assurance purposes. To evaluate the practical feasibility and performance of the QualiPilot tool, a series of benchmark tests were conducted using three computer systems, with a particular focus on devices with high, moderate and low hardware configurations. All tests were performed under Windows 11 and Windows 10, using the local model “llama-2-7b-chat-hf-q4_k_m.gguf” [26]. The primary aim of this study was to assess whether the tool can generate meaningful and structured qualitative analysis results under realistic conditions, without relying on cloud-based infrastructures or dedicated GPUs. For this purpose, the study by Ben-Shabat et al. (2024) was selected, as its results, including the full interview transcripts, were published only after the release of the LLM model [28]. This timing ensured that the model could not have been exposed to these findings during its training. In this study, the results generated by QualiPilot were compared with the original study findings by the authors as an initial step. A comprehensive qualitative evaluation of the QualiPilot outputs will be carried out in a subsequent, dedicated evaluation study involving independent reviewers. Thus, the findings reported here are intended solely as a preliminary appraisal to inform prompt design and model parameter selection, rather than as a full evaluation. Results All computers handled the core analytic steps of the workflow well, including initial text analysis and the development of an inductive coding system. The maximum input size was capped at 4,096 tokens, reflecting the chosen model limitations. The number of threads was set in line with the available hardware. To illustrate the influence of hardware performance, tests were conducted on three different systems: a dated notebook with an Intel i5-4210U processor (1.7 GHz, 4 threads), a modern ultrabook with an Intel i5-1335U (1.3 GHz, 12 threads), and a high-end workstation equipped with an AMD Ryzen AI MAY + 395 (3.0 GHz, 30 threads). For reference, 4,096 tokens correspond to approximately 3,000 words of English text. On recent hardware, such as the AMD Ryzen system, the average output speed was around 7 to 9 tokens per second, resulting in a processing time of about 8 minutes for an input of 3,000 words. In contrast, on the older notebook with the i5-4210U, output speed decreased to 0.7 to 0.8 tokens per second, leading to a processing time of up to 140 minutes for the same input. The workflow and performance assessments were conducted using real interview data to ensure practical relevance and validity. This demonstrates that, while full transcripts or interview excerpts with up to 3,000 words can be processed in a single step, response times are heavily dependent on the available hardware. All relevant hardware specifications and measured performance values are provided in Table 2 . Table 2 Performance Benchmarks for QualiPilot on Different Hardware Setups PC RAM (GB) Processor / GPU Threads tokens per second Max RAM Used (GB) Windows Time needed to process a 3,000 word English-language transcript 1 8 Intel i5-4210U 1.70 GHz 4 0.7–0.8 4.5 10 Pro 85–97 min 2 16 Intel i5-1335U, 1.30 GHz 12 1.5–2.5 5.4 11 Pro 24–46 min 3 128 AMD Ryzen AI MAY + 395, 3.00 GHz 30 7.0–9.0 9.3 11 Pro 8–10 min Notes : All analyses were conducted with an input limit of 4,096 tokens, set according to the model's and system's capabilities; No software errors, crashes, or instability were observed. Results indicate that, while the QualiPilot tool can run on standard office hardware, the processing speed is significantly slower compared to GPU-accelerated or online/cloud-based solutions. Despite longer runtimes, especially on lower-end systems, all analytic steps produced results that were comparable to the findings of the original study. The system- and task-specific prompts consistently resulted in clearly formatted category systems and code lists, with short descriptions for each subcategory. No qualitative differences in the logical consistency or structure of the results were observed between the different hardware setups. RAM consumption during inference remained well below the physical limits of the tested hardware, confirming the tool's suitability for devices with moderate specifications. Throughout all tests, the application remained stable, responsive, and error-free, even when operating near its token or memory limits. For the initial text analysis, the outputs addressed the research objectives and reflected the main topics identified in the original work, although some responses were more elaborated or presented information in a slightly different way. In the inductive coding system step, the tool generated category frameworks that were broadly consistent with the main categories and subcategories of the original study, while sometimes providing additional detail or alternative structuring. For the coding of individual segments, the tool produced outputs that covered relevant aspects of the content, though not always using identical language or emphasis. However, due to programming and hardware limitations, this step exhibited slower processing speeds compared to the other steps. On standard office hardware, the time required to produce coded outputs is a limiting factor, particularly for projects involving larger datasets or those under tight time constraints. The code summarization results were clear and generally suitable for synthesis and reporting purposes. It should be noted that all comparisons were made against the original study results and assessed by the authors themselves. While this comparison showed substantial overlap, a comprehensive evaluation by independent reviewers was not conducted within the scope of this study. A recurring issue is the context window and token limit imposed by the underlying LLM architecture. With a maximum of 4,096 tokens per analytic step, longer texts or interview transcripts require segmentation into smaller portions. While this is manageable in practice, it disrupted the holistic flow of analysis and requires additional manual coordination. In summary, while the offline local implementation is considerably slower than current online or GPU-based alternatives, QualiPilot produced analytic outputs that were consistent in structure and generally well-aligned with the original study findings. The presentation of results always followed the same pattern, which was straightforward to read and interpret. Conclusion This study introduces and systematically evaluates QualiPilot, a locally deployable, fully offline AI-assisted tool for qualitative content analysis. The findings underscore that QualiPilot offers a practical entry point for researchers and institutions seeking to explore the potential of large language models in qualitative research, while adhering to rigorous standards for data privacy, methodological transparency, and ethical research conduct. By operating independently of commercial cloud services and relying solely on local hardware, QualiPilot offers an accessible, low-barrier option for AI-supported qualitative analysis, making it especially suitable for sensitive or confidential research contexts. A key strength of QualiPilot is its analytic performance even on standard computing devices. The tool consistently delivers relevant and well-structured outputs for initial text analysis, inductive coding, text coding, and code summarization tasks. All analytic steps are fully transparent and reproducible, with model parameters and system prompts documented for each task. However, the present results also show that these advantages come at the cost of processing speed. Compared to cloud-based or GPU-accelerated solutions, locally deployed LLMs—especially when run on basic hardware—require substantially more time to generate analytic outputs. For researchers with limited technical infrastructure, this may entail a significant trade-off between turnaround time and data security. Nevertheless, with access to more powerful hardware, processing times can be markedly improved, and even the basic version of QualiPilot delivers good results within reasonable time frames for most typical qualitative projects. Overall, QualiPilot is not intended to replace traditional manual qualitative analysis or the critical interpretive role of the researcher. Rather, its most promising application is as a supplementary tool, offering a "second opinion," enabling triangulation, or providing rapid initial analytic frameworks that can then be further developed and refined by researchers. It could also be used as a validation instrument for the analyzed data. While qualitative analyses are often conducted collaboratively, they typically involve similar groups of researchers, and systematic external quality assurance is rarely feasible due to limited resources. Once experience has been built up, more complex local LLMs could also be used in other applications with or without Python, which offer faster performance but demand more technical expertise and configuration effort. The clear structure and high interpretability of the QualiPilot outputs make it especially useful for researchers seeking to experiment with AI-supported methods in a controlled, privacy-compliant environment. Limitations Despite its strengths, QualiPilot has several notable limitations. The current implementation is deliberately designed with a minimalist, basic programming approach to maximize accessibility for the academic community. As a result, some functionalities, interface elements, and user experience aspects may be less advanced than those found in established commercial CAQDAS platforms. The tool is currently optimized for smaller to medium datasets due to the model’s context window/token limit, limiting the volume of text that can be processed per analytic cycle. Longer transcripts require segmentation and may result in fragmented analytic processes. Future Directions and Further Applications QualiPilot is freely available for non-commercial research use, and all interested users are encouraged to download, test, and provide feedback. QualiPilot will continue to evolve based on feedback from its user community. We actively invite users to share their feedback, report bugs, and suggest new features. All input can be directed to the corresponding author and will be considered in shaping future updates. Planned enhancements include the integration of larger and more advanced language models for high-performance systems, further optimization of memory and processing routines, and expanded options for prompt customization and analytic workflows. Additionally, the tool is expected to benefit from ongoing advances in local LLM technology, including model quantization and edge computing. Beyond its immediate application in research, QualiPilot can also serve as an educational resource for teaching AI-assisted qualitative methods and as a privacy-compliant alternative for initial pilot analyses of sensitive data. In summary, QualiPilot offers a novel, robust, and community-oriented platform for offline AI-assisted qualitative research. While inevitable trade-offs exist, particularly in terms of speed and advanced feature set, its strengths in transparency, data security, and analytic clarity make it a valuable addition to the methodological repertoire of qualitative researchers. Ongoing user engagement and technological evolution will further enhance its utility and impact within the broader field of qualitative inquiry. Declarations Ethics approval and consent to participate Ethics approval for the entire project was not required. Clinical trial number Not Applicable. Consent for publication Not applicable Availability of data and materials The tool is available as a download from: Kempny, C. (2025). QualiPilot – an offline AI tool for qualitative content analysis (1.0). Zenodo. https://doi.org/10.5281/zenodo.15879818 The outputs generated by the tool are provided in the Supplementary Material. Competing interests The authors declare that they have no competing interests. Funding This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Authors' contributions CK conceptualized, designed, and conducted the study. CK developed the software and performed all analyses. Validation and testing were carried out by CK, KA, YYA and PB. Data extraction and curation were performed by CK. CK drafted the manuscript, with review and editing contributions from KA, YYA and PB. Supervision was provided by PB. All authors reviewed and approved the final version of the manuscript. Acknowledgments Not applicable References Summers Holtrop J, Williams J, Connelly L, Perreault L. AI-Drive Targeted Qualitative Analysis: A Useful New Method for Primary Care Research. In: Qualitative research. American Academy of Family Physicians; 2024. p. 6343. Gao J, Choo KTW, Cao J, Lee RKW, Perrault S. CoAIcoder: Examining the Effectiveness of AI-assisted Human-to-Human Collaboration in Qualitative Analysis. 2023. Christou P. A Critical Perspective Over Whether and How to Acknowledge the Use of Artificial Intelligence (AI) in Qualitative Studies. TQR. 2023. https://doi.org/10.46743/2160-3715/2023.6407. Ashwin J, Chhabra A, Rao V. Using Large Language Models for Qualitative Analysis can Introduce Serious Bias. 2023. Roberts J, Baker M, Andrew J. Artificial intelligence and qualitative research: The promise and perils of large language model (LLM) ‘assistance.’ Critical Perspectives on Accounting. 2024;99:102722. Kwon H, Min S, Lee S. How to Better Translate Participant Quotes Using LLMs: Exploring Practices and Challenges of Non-Native English Researchers. In: Proceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems. Yokohama Japan: ACM; 2025. p. 1–11. Balt E, Salmi S, Bhulai S, Vrinzen S, Eikelenboom M, Gilissen R, et al. Deductively coding psychosocial autopsy interview data using a few-shot learning large language model. Front Public Health. 2025;13. Kapania S, Agnew W, Eslami M, Heidari H, Fox SE. Simulacrum of Stories: Examining Large Language Models as Qualitative Research Participants. In: Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. Yokohama Japan: ACM; 2025. p. 1–17. VERBI Software – Consult -Sozialforschung GmbH. MAXQDA 24. 2024. ATLAS.ti Scientific Software Development GmbH. ATLAS.ti. 2025. European Data Protection Board. Study on the secondary use of personal data in the context of scientific research. 2020. Da Silva J. Producing ‘good enough’ automated transcripts securely: Extending Bokhove and Downey (2018) to address security concerns. Methodological Innovations. 2021;14. Sponholz J, Weilinghoff A, Schopf J. Halving transcription time: A fast, user-friendly and GDPR-compliant workflow to create AI-assisted transcripts for content analysis. 2025. forschungsdaten.info. Datenübertragung. Forschungsdaten.info. 2023. https://forschungsdaten.info/themen/organisieren-und-aufbereiten/datenuebertragung/. Accessed 14 May 2025. MAXQDA - Distribution by VERBI GmbH. AI Assist. FAQ. 2025. https://help.maxqda.com/en/support/solutions/folders/80000718353. ATLAS.ti. KI Datensicherheit und Datenschutz bei ATLAS.ti. 2025. https://atlasti.com/de/ai-datensicherheit-und-datenschutz. Sclar M, Choi Y, Tsvetkov Y, Suhr A. Quantifying Language Models’ Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting. 2023. Xu L, Zhang J, Li B, Wang J, Chen S, Zhao WX, et al. Tapping the Potential of Large Language Models as Recommender Systems: A Comprehensive Framework and Empirical Analysis. ACM Trans Knowl Discov Data. 2025;19:1–51. He J, Rungta M, Koleczek D, Sekhon A, Wang FX, Hasan S. Does Prompt Formatting Have Any Impact on LLM Performance? 2024. Zamfirescu-Pereira JD, Wong RY, Hartmann B, Yang Q. Why Johnny Can’t Prompt: How Non-AI Experts Try (and Fail) to Design LLM Prompts. In: Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. Hamburg Germany: ACM; 2023. p. 1–21. Fehring L, Frings J, Rust P, Kempny C, Thürmann P, Meister S. Reporting of Qualitative Research Using Large Language Models (COREQ + LLM): Protocol for an Extension of the Consolidated Criteria for Reporting Qualitative Research Guideline (Preprint). 2025. Perron BE, Luan H, Victor BG, Hiltz-Perron O, Ryan J. Moving Beyond ChatGPT: Local Large Language Models (LLMs) and the Secure Analysis of Confidential Unstructured Text Data in Social Work Research. Research on Social Work Practice. 2024. https://doi.org/10.1177/10497315241280686. Meyer T, Baker C, Keefe J. Enhancing Thematic Analysis with Local LLMs: A Scientific Evaluation of Prompt Engineering Techniques. In: AHFE International. AHFE International; 2025. Bennis I, Mouwafaq S. Advancing AI-driven thematic analysis in qualitative research: a comparative study of nine generative models on Cutaneous Leishmaniasis data. BMC Med Inform Decis Mak. 2025;25. Wheldon C, McKee R. AI-empowered Qualitative Data Analysis: Training Future Public Health Researchers. CH. 2025;6. TheBloke. Llama 2 7B Chat - GGUF. 2023. Holtzman A, Buys J, Du L, Forbes M, Choi Y. The Curious Case of Neural Text Degeneration. 2019. Ben-Shabat I, Lindvall K, Salzer J. Exploring strategies for management of in-hospital stroke in Sweden: A qualitative study. PLoS ONE. 2024;19:e0313765. Additional Declarations No competing interests reported. Supplementary Files SuplementMaterial.docx 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-7122121","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":485521561,"identity":"c2e9becf-c4aa-496a-a09d-59d381f36eee","order_by":0,"name":"Christian Kempny","email":"data:image/png;base64,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","orcid":"","institution":"Witten/Herdecke University","correspondingAuthor":true,"prefix":"","firstName":"Christian","middleName":"","lastName":"Kempny","suffix":""},{"id":485521562,"identity":"7fd915d1-9afb-4573-9598-e54a39430640","order_by":1,"name":"Kübra Annac","email":"","orcid":"","institution":"Witten/Herdecke University","correspondingAuthor":false,"prefix":"","firstName":"Kübra","middleName":"","lastName":"Annac","suffix":""},{"id":485521563,"identity":"59df0d44-5ea4-4452-bdc8-f98e5b4d0fc7","order_by":2,"name":"Yuece Yilmaz-Aslan","email":"","orcid":"","institution":"Witten/Herdecke University","correspondingAuthor":false,"prefix":"","firstName":"Yuece","middleName":"","lastName":"Yilmaz-Aslan","suffix":""},{"id":485521564,"identity":"4fe4a9e6-1b1d-476a-848e-e2d310de332d","order_by":3,"name":"Patrick Brzoska","email":"","orcid":"","institution":"Witten/Herdecke University","correspondingAuthor":false,"prefix":"","firstName":"Patrick","middleName":"","lastName":"Brzoska","suffix":""}],"badges":[],"createdAt":"2025-07-14 14:23:20","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7122121/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7122121/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":86958905,"identity":"94fb2062-dfbf-447b-803a-8119efd0d086","added_by":"auto","created_at":"2025-07-17 15:38:21","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":572153,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7122121/v1/d9659e4e-9c2e-49ef-ab5b-3064bcf74330.pdf"},{"id":86839443,"identity":"8e487e4a-744d-4336-940d-577bc54f5528","added_by":"auto","created_at":"2025-07-16 07:46:56","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":39110,"visible":true,"origin":"","legend":"","description":"","filename":"SuplementMaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-7122121/v1/4a3dd529f55dbd1750fcc909.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"QualiPilot – an offline AI tool for qualitative content analysis","fulltext":[{"header":"Background","content":"\u003cp\u003eInterest in AI-assisted qualitative content analysis has increased in recent years, with researchers exploring a range of approaches from traditional enhancements of established qualitative methods to the development of entirely new, AI-driven analytical frameworks [1–3]. The use of LLMs in qualitative research offers significant opportunities but also entails substantial risks that span methodological, ethical, and epistemological domains. Key concerns include systematic bias introduced by LLMs, which can lead to distorted or misleading findings, particularly for sensitive topics or marginalized groups [4–6]. LLMs are prone to generating “hallucinations” (plausible but factually incorrect or fabricated content), often presented with confidence, increasing the risk of uncritical acceptance by researchers [5, 7]. Moreover, LLM-generated outputs may lack contextual depth and overlook important cultural or individual nuances, thereby affecting the credibility and richness of qualitative analysis [6, 8]. To mitigate these risks, it is essential to ensure critical human oversight, maintain transparency in LLM use and prompt design, and uphold strict ethical and data protection standards.\u003c/p\u003e\u003cp\u003eStill, despite these shortcomings, LLMs have been shown to serve as valuable research partners by providing complementary insights that support and enrich the research process [13–15]. Additionally, LLMs can reveal patterns and connections often overlooked in manual review, particularly when researchers “converse” with their materials by asking targeted questions, probing for contextual insights, and refining theoretical connections [16, 17].\u003c/p\u003e\u003cp\u003eHowever, a persistent challenge in exploiting that potential is that most practical solutions currently rely on cloud-based infrastructures, such as MAXQDA AI Assist [9], ATLAS.ti [10] and various chatbot-based platforms like OpenAI or Meta. While these tools make powerful language models accessible for qualitative analysis, they raise substantial concerns regarding data protection, transparency, and compliance with research ethics. Sensitive qualitative data, particularly interview transcripts and personal narratives, are often subject to stringent legal and ethical requirements that mandate the researcher’s direct control over the data and limits its transfer to external servers [11–14]. Moreover, a significant limitation of existing online solutions is their lack of transparency and reproducibility. Many cloud-based AI tools do not disclose critical information such as which large language model (LLM) and version is being used, which parameters and prompts are applied, or how system prompts influence the analytic output [15, 16]. However, prompting strategies as well as the version and configuration of an LLM can significantly influence analytic outcomes [17–20]. Such details are rarely disclosed in the reporting of qualitative studies that utilize AI support [21]. This opacity complicates both the reporting and replication of research findings, raising concerns about scientific validity, methodological rigor, and ethical and legal accountability regarding data transparency.\u003c/p\u003e\u003cp\u003eThe integration of offline, locally deployed AI models into qualitative data analysis tools offers a compelling response to these limitations. Offline large language models (Local LLMs) can be run directly on researchers’ computers, enabling complete data sovereignty, maximum transparency, and reproducibility [22]. In addition, all analytic steps, model parameters, and prompt designs can be fully documented and controlled by the research team, supporting comprehensive, transparent reporting and robust reproducibility of findings [23]. Methodologically, the integration of local LLMs enables several new opportunities in qualitative analysis. From a quality assurance perspective, local AI systems enable rigorous validation processes.\u003c/p\u003e\u003cp\u003eGiven evidence showing that AI-assisted qualitative analysis can support researchers [2, 24, 25], the availability of local LLMs and structured reporting of all information makes these analyses reproducible without dependence on online solutions that do not disclose key information or on tools that modify the LLM through updates.\u003c/p\u003e\u003cp\u003eFinally, from a legal and ethical standpoint, local AI solutions simplify compliance with data protection regulations and institutional ethics requirements. Because no data leaves the researcher’s environment, and all processing steps can be precisely documented, informed consent and ethics approvals are easier to secure and substantiate for review boards and funding agencies.\u003c/p\u003e\u003cp\u003eWith the rise of edge computing and quantized models, powerful LLMs are becoming increasingly accessible on standard hardware. A key advantage of local deployment is that the LLM remains permanently available in a fixed version because updates occur only with explicit model releases. This provides consistent access and reproducible outputs, which is not guaranteed when using cloud-based services where automatic updates can unpredictably change model behavior and affect performance. However, running large language models locally also introduces practical challenges. Users may face barriers such as insufficient administrative rights or complicated installation processes, particularly within institutional environments. In addition, limited familiarity with technical concepts such as model selection, prompt engineering, or system prompts can further hinder effective use in research settings. Designing appropriate prompts tailored to the specific LLM is therefore essential to enable meaningful and reliable application.\u003c/p\u003e\u003cp\u003eConsidering these challenges, this study aims to develop an application-friendly, local LLM solution (‘QualiPilot‘) tailored explicitly for qualitative research. The goal is to provide a tool that not only functions entirely on local hardware but also prioritizes user accessibility through an intuitive interface and the ability to run directly from a USB flash drive without installation. By focusing on comprehensive and transparent reporting of all analytic parameters and system prompts, the tool enables full reproducibility and methodological transparency. Importantly, this solution is designed to be freely available for use in academic and research contexts, thereby lowering the entry barriers for researchers who may lack access to commercial or cloud-based AI services or are hesitant to use them.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eThe QualiPilot application was developed as a stand-alone, privacy-compliant, and fully offline desktop tool for AI-assisted qualitative content analysis. In contrast to traditional CAQDAS platforms, QualiPilot operates independently as a chat-based application. Its design enables researchers to perform key steps of inductive qualitative analysis, including initial text engagement, inductive category development, coding, and code summarization, by interacting with a local LLM.\u003c/p\u003e\u003cp\u003eThe technical implementation of QualiPilot prioritizes accessibility and transparency. The tool is written in Python and uses the llama_cpp library [26] to interface with a small llama-2 model, which can be stored and executed locally with minimal RAM requirements. All computations and text processing are performed entirely on the user's computer, ensuring that no data is transmitted to external servers. The application is optimized for standard hardware, requiring a minimum of 8 GB of RAM and 4 threads, making it broadly accessible to most users. For high-performance systems, a larger and more advanced model variant will be available, though this exceeds the resource capacity of many typical PCs.\u003c/p\u003e\u003cp\u003eA key feature of QualiPilot is its ability to apply tailored parameter settings for each analytical step, dynamically adjusting these settings to suit both the analytical task at hand and the capabilities of the chosen LLM model. For each step—initial text engagement, category development, coding, and code summarization—specific model settings are used to optimize performance and analytic quality. The parameter settings for each analytic task are shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. The user interface makes all model and parameter information transparent, encouraging the documentation of these settings for reproducibility.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eLLM Model Parameters Used in QualiPilot for Different Analytic Tasks\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"8\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAnalytic Task\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eTemperature\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eTop-k\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTop-p\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eMin-p\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eRepeat Penalty\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eMax Answer Tokens\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eMax Input Tokens\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eInitial text engagement / Memoing\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e1.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e1024\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e4096\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eInductive category development\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e1.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e2048\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e4096\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eCoding (individual segments)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e1.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e512\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e4096\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eCode summarization\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e1.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e2048\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e4096\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"8\"\u003e\u003cb\u003eNote.\u003c/b\u003e Each parameter controls a specific aspect of text generation. Temperature (usually 0.1–2.0) sets how creative or random the answers are. Higher temperature values lead to more creative, lower values to more predictable responses. Top-k (1–100) means the model will only choose from the k most likely next words instead of considering all possibilities. Top-p (0–1) works a bit differently: the model considers the most likely next words until their total probability adds up to p, and then picks from this group. Min-p (0–1) filters out words that are too unlikely. Repeat Penalty (usually 1.0–2.0) helps avoid repeating the same words or phrases. Max Answer Tokens sets how long the answer can be, and Max Input Tokens limits how much text can be input at once [27].\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eAll analytic tasks are guided by specialized system prompts tailored for qualitative research. The workflow is designed to enable users to conduct initial reading and develop a unified inductive category system, assign codes, and create code summaries, with each step’s outputs being fully documented. All analytical steps and settings are listed and can be exported for reporting and quality assurance purposes.\u003c/p\u003e\u003cp\u003eTo evaluate the practical feasibility and performance of the QualiPilot tool, a series of benchmark tests were conducted using three computer systems, with a particular focus on devices with high, moderate and low hardware configurations. All tests were performed under Windows 11 and Windows 10, using the local model “llama-2-7b-chat-hf-q4_k_m.gguf” [26]. The primary aim of this study was to assess whether the tool can generate meaningful and structured qualitative analysis results under realistic conditions, without relying on cloud-based infrastructures or dedicated GPUs.\u003c/p\u003e\u003cp\u003eFor this purpose, the study by Ben-Shabat et al. (2024) was selected, as its results, including the full interview transcripts, were published only after the release of the LLM model [28]. This timing ensured that the model could not have been exposed to these findings during its training. In this study, the results generated by QualiPilot were compared with the original study findings by the authors as an initial step. A comprehensive qualitative evaluation of the QualiPilot outputs will be carried out in a subsequent, dedicated evaluation study involving independent reviewers. Thus, the findings reported here are intended solely as a preliminary appraisal to inform prompt design and model parameter selection, rather than as a full evaluation.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eAll computers handled the core analytic steps of the workflow well, including initial text analysis and the development of an inductive coding system. The maximum input size was capped at 4,096 tokens, reflecting the chosen model limitations. The number of threads was set in line with the available hardware. To illustrate the influence of hardware performance, tests were conducted on three different systems: a dated notebook with an Intel i5-4210U processor (1.7 GHz, 4 threads), a modern ultrabook with an Intel i5-1335U (1.3 GHz, 12 threads), and a high-end workstation equipped with an AMD Ryzen AI MAY\u0026thinsp;+\u0026thinsp;395 (3.0 GHz, 30 threads).\u003c/p\u003e\u003cp\u003eFor reference, 4,096 tokens correspond to approximately 3,000 words of English text. On recent hardware, such as the AMD Ryzen system, the average output speed was around 7 to 9 tokens per second, resulting in a processing time of about 8 minutes for an input of 3,000 words. In contrast, on the older notebook with the i5-4210U, output speed decreased to 0.7 to 0.8 tokens per second, leading to a processing time of up to 140 minutes for the same input. The workflow and performance assessments were conducted using real interview data to ensure practical relevance and validity.\u003c/p\u003e\u003cp\u003eThis demonstrates that, while full transcripts or interview excerpts with up to 3,000 words can be processed in a single step, response times are heavily dependent on the available hardware. All relevant hardware specifications and measured performance values are provided in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePerformance Benchmarks for QualiPilot on Different Hardware Setups\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"8\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRAM (GB)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eProcessor / GPU\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eThreads\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003etokens per second\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eMax RAM Used (GB)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eWindows\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eTime needed to process a 3,000 word English-language transcript\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eIntel i5-4210U 1.70 GHz\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.7\u0026ndash;0.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e4.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e10 Pro\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e85\u0026ndash;97 min\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eIntel i5-1335U, 1.30 GHz\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e1.5\u0026ndash;2.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e5.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e11 Pro\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e24\u0026ndash;46 min\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e128\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAMD Ryzen AI MAY\u0026thinsp;+\u0026thinsp;395, 3.00 GHz\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e30\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e7.0\u0026ndash;9.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e9.3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e11 Pro\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e8\u0026ndash;10 min\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"8\"\u003e\u003cb\u003eNotes\u003c/b\u003e: All analyses were conducted with an input limit of 4,096 tokens, set according to the model's and system's capabilities; No software errors, crashes, or instability were observed.\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eResults indicate that, while the QualiPilot tool can run on standard office hardware, the processing speed is significantly slower compared to GPU-accelerated or online/cloud-based solutions. Despite longer runtimes, especially on lower-end systems, all analytic steps produced results that were comparable to the findings of the original study. The system- and task-specific prompts consistently resulted in clearly formatted category systems and code lists, with short descriptions for each subcategory. No qualitative differences in the logical consistency or structure of the results were observed between the different hardware setups.\u003c/p\u003e\u003cp\u003eRAM consumption during inference remained well below the physical limits of the tested hardware, confirming the tool's suitability for devices with moderate specifications. Throughout all tests, the application remained stable, responsive, and error-free, even when operating near its token or memory limits.\u003c/p\u003e\u003cp\u003eFor the initial text analysis, the outputs addressed the research objectives and reflected the main topics identified in the original work, although some responses were more elaborated or presented information in a slightly different way. In the inductive coding system step, the tool generated category frameworks that were broadly consistent with the main categories and subcategories of the original study, while sometimes providing additional detail or alternative structuring. For the coding of individual segments, the tool produced outputs that covered relevant aspects of the content, though not always using identical language or emphasis. However, due to programming and hardware limitations, this step exhibited slower processing speeds compared to the other steps. On standard office hardware, the time required to produce coded outputs is a limiting factor, particularly for projects involving larger datasets or those under tight time constraints. The code summarization results were clear and generally suitable for synthesis and reporting purposes. It should be noted that all comparisons were made against the original study results and assessed by the authors themselves. While this comparison showed substantial overlap, a comprehensive evaluation by independent reviewers was not conducted within the scope of this study.\u003c/p\u003e\u003cp\u003eA recurring issue is the context window and token limit imposed by the underlying LLM architecture. With a maximum of 4,096 tokens per analytic step, longer texts or interview transcripts require segmentation into smaller portions. While this is manageable in practice, it disrupted the holistic flow of analysis and requires additional manual coordination.\u003c/p\u003e\u003cp\u003eIn summary, while the offline local implementation is considerably slower than current online or GPU-based alternatives, QualiPilot produced analytic outputs that were consistent in structure and generally well-aligned with the original study findings. The presentation of results always followed the same pattern, which was straightforward to read and interpret.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study introduces and systematically evaluates QualiPilot, a locally deployable, fully offline AI-assisted tool for qualitative content analysis. The findings underscore that QualiPilot offers a practical entry point for researchers and institutions seeking to explore the potential of large language models in qualitative research, while adhering to rigorous standards for data privacy, methodological transparency, and ethical research conduct. By operating independently of commercial cloud services and relying solely on local hardware, QualiPilot offers an accessible, low-barrier option for AI-supported qualitative analysis, making it especially suitable for sensitive or confidential research contexts.\u003c/p\u003e\u003cp\u003eA key strength of QualiPilot is its analytic performance even on standard computing devices. The tool consistently delivers relevant and well-structured outputs for initial text analysis, inductive coding, text coding, and code summarization tasks. All analytic steps are fully transparent and reproducible, with model parameters and system prompts documented for each task.\u003c/p\u003e\u003cp\u003eHowever, the present results also show that these advantages come at the cost of processing speed. Compared to cloud-based or GPU-accelerated solutions, locally deployed LLMs\u0026mdash;especially when run on basic hardware\u0026mdash;require substantially more time to generate analytic outputs. For researchers with limited technical infrastructure, this may entail a significant trade-off between turnaround time and data security. Nevertheless, with access to more powerful hardware, processing times can be markedly improved, and even the basic version of QualiPilot delivers good results within reasonable time frames for most typical qualitative projects.\u003c/p\u003e\u003cp\u003eOverall, QualiPilot is not intended to replace traditional manual qualitative analysis or the critical interpretive role of the researcher. Rather, its most promising application is as a supplementary tool, offering a \"second opinion,\" enabling triangulation, or providing rapid initial analytic frameworks that can then be further developed and refined by researchers. It could also be used as a validation instrument for the analyzed data. While qualitative analyses are often conducted collaboratively, they typically involve similar groups of researchers, and systematic external quality assurance is rarely feasible due to limited resources.\u003c/p\u003e\u003cp\u003eOnce experience has been built up, more complex local LLMs could also be used in other applications with or without Python, which offer faster performance but demand more technical expertise and configuration effort. The clear structure and high interpretability of the QualiPilot outputs make it especially useful for researchers seeking to experiment with AI-supported methods in a controlled, privacy-compliant environment.\u003c/p\u003e\u003cp\u003eLimitations\u003c/p\u003e\u003cp\u003eDespite its strengths, QualiPilot has several notable limitations. The current implementation is deliberately designed with a minimalist, basic programming approach to maximize accessibility for the academic community. As a result, some functionalities, interface elements, and user experience aspects may be less advanced than those found in established commercial CAQDAS platforms. The tool is currently optimized for smaller to medium datasets due to the model\u0026rsquo;s context window/token limit, limiting the volume of text that can be processed per analytic cycle. Longer transcripts require segmentation and may result in fragmented analytic processes.\u003c/p\u003e\u003cp\u003eFuture Directions and Further Applications\u003c/p\u003e\u003cp\u003eQualiPilot is freely available for non-commercial research use, and all interested users are encouraged to download, test, and provide feedback. QualiPilot will continue to evolve based on feedback from its user community. We actively invite users to share their feedback, report bugs, and suggest new features. All input can be directed to the corresponding author and will be considered in shaping future updates.\u003c/p\u003e\u003cp\u003ePlanned enhancements include the integration of larger and more advanced language models for high-performance systems, further optimization of memory and processing routines, and expanded options for prompt customization and analytic workflows. Additionally, the tool is expected to benefit from ongoing advances in local LLM technology, including model quantization and edge computing.\u003c/p\u003e\u003cp\u003eBeyond its immediate application in research, QualiPilot can also serve as an educational resource for teaching AI-assisted qualitative methods and as a privacy-compliant alternative for initial pilot analyses of sensitive data.\u003c/p\u003e\u003cp\u003eIn summary, QualiPilot offers a novel, robust, and community-oriented platform for offline AI-assisted qualitative research. While inevitable trade-offs exist, particularly in terms of speed and advanced feature set, its strengths in transparency, data security, and analytic clarity make it a valuable addition to the methodological repertoire of qualitative researchers. Ongoing user engagement and technological evolution will further enhance its utility and impact within the broader field of qualitative inquiry.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eEthics approval for the entire project was not required.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical trial number\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot Applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eThe tool is available as a download from:\u0026nbsp;\u003c/strong\u003eKempny, C. (2025). QualiPilot \u0026ndash; an offline AI tool for qualitative content analysis (1.0). Zenodo. https://doi.org/10.5281/zenodo.15879818\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe outputs generated by the tool are provided in the Supplementary Material.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026apos; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCK conceptualized, designed, and conducted the study. CK developed the software and performed all analyses. Validation and testing were carried out by CK, KA, YYA and PB. Data extraction and curation were performed by CK. CK drafted the manuscript, with review and editing contributions from KA, YYA and PB. Supervision was provided by PB. All authors reviewed and approved the final version of the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003e\u003cspan\u003eSummers Holtrop J, Williams J, Connelly L, Perreault L. AI-Drive Targeted Qualitative Analysis: A Useful New Method for Primary Care Research. In: Qualitative research. American Academy of Family Physicians; 2024. p. 6343.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eGao J, Choo KTW, Cao J, Lee RKW, Perrault S. CoAIcoder: Examining the Effectiveness of AI-assisted Human-to-Human Collaboration in Qualitative Analysis. 2023.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eChristou P. A Critical Perspective Over Whether and How to Acknowledge the Use of Artificial Intelligence (AI) in Qualitative Studies. TQR. 2023. https://doi.org/10.46743/2160-3715/2023.6407.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eAshwin J, Chhabra A, Rao V. Using Large Language Models for Qualitative Analysis can Introduce Serious Bias. 2023.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eRoberts J, Baker M, Andrew J. Artificial intelligence and qualitative research: The promise and perils of large language model (LLM) \u0026lsquo;assistance.\u0026rsquo; Critical Perspectives on Accounting. 2024;99:102722.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eKwon H, Min S, Lee S. How to Better Translate Participant Quotes Using LLMs: Exploring Practices and Challenges of Non-Native English Researchers. In: Proceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems. Yokohama Japan: ACM; 2025. p. 1\u0026ndash;11.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eBalt E, Salmi S, Bhulai S, Vrinzen S, Eikelenboom M, Gilissen R, et al. Deductively coding psychosocial autopsy interview data using a few-shot learning large language model. Front Public Health. 2025;13.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eKapania S, Agnew W, Eslami M, Heidari H, Fox SE. Simulacrum of Stories: Examining Large Language Models as Qualitative Research Participants. In: Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. Yokohama Japan: ACM; 2025. p. 1\u0026ndash;17.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eVERBI Software \u0026ndash; Consult -Sozialforschung GmbH. MAXQDA 24. 2024.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eATLAS.ti Scientific Software Development GmbH. ATLAS.ti. 2025.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eEuropean Data Protection Board. Study on the secondary use of personal data in the context of scientific research. 2020.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eDa Silva J. Producing \u0026lsquo;good enough\u0026rsquo; automated transcripts securely: Extending Bokhove and Downey (2018) to address security concerns. Methodological Innovations. 2021;14.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eSponholz J, Weilinghoff A, Schopf J. Halving transcription time: A fast, user-friendly and GDPR-compliant workflow to create AI-assisted transcripts for content analysis. 2025.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eforschungsdaten.info. Daten\u0026uuml;bertragung. Forschungsdaten.info. 2023. https://forschungsdaten.info/themen/organisieren-und-aufbereiten/datenuebertragung/. Accessed 14 May 2025.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eMAXQDA - Distribution by VERBI GmbH. AI Assist. FAQ. 2025. https://help.maxqda.com/en/support/solutions/folders/80000718353.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eATLAS.ti. KI Datensicherheit und Datenschutz bei ATLAS.ti. 2025. https://atlasti.com/de/ai-datensicherheit-und-datenschutz.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eSclar M, Choi Y, Tsvetkov Y, Suhr A. Quantifying Language Models\u0026rsquo; Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting. 2023.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eXu L, Zhang J, Li B, Wang J, Chen S, Zhao WX, et al. Tapping the Potential of Large Language Models as Recommender Systems: A Comprehensive Framework and Empirical Analysis. ACM Trans Knowl Discov Data. 2025;19:1\u0026ndash;51.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eHe J, Rungta M, Koleczek D, Sekhon A, Wang FX, Hasan S. Does Prompt Formatting Have Any Impact on LLM Performance? 2024.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eZamfirescu-Pereira JD, Wong RY, Hartmann B, Yang Q. Why Johnny Can\u0026rsquo;t Prompt: How Non-AI Experts Try (and Fail) to Design LLM Prompts. In: Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. Hamburg Germany: ACM; 2023. p. 1\u0026ndash;21.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eFehring L, Frings J, Rust P, Kempny C, Th\u0026uuml;rmann P, Meister S. Reporting of Qualitative Research Using Large Language Models (COREQ\u0026thinsp;+\u0026thinsp;LLM): Protocol for an Extension of the Consolidated Criteria for Reporting Qualitative Research Guideline (Preprint). 2025.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003ePerron BE, Luan H, Victor BG, Hiltz-Perron O, Ryan J. Moving Beyond ChatGPT: Local Large Language Models (LLMs) and the Secure Analysis of Confidential Unstructured Text Data in Social Work Research. Research on Social Work Practice. 2024. https://doi.org/10.1177/10497315241280686.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eMeyer T, Baker C, Keefe J. Enhancing Thematic Analysis with Local LLMs: A Scientific Evaluation of Prompt Engineering Techniques. In: AHFE International. AHFE International; 2025.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eBennis I, Mouwafaq S. Advancing AI-driven thematic analysis in qualitative research: a comparative study of nine generative models on Cutaneous Leishmaniasis data. BMC Med Inform Decis Mak. 2025;25.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eWheldon C, McKee R. AI-empowered Qualitative Data Analysis: Training Future Public Health Researchers. CH. 2025;6.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eTheBloke. Llama 2 7B Chat - GGUF. 2023.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eHoltzman A, Buys J, Du L, Forbes M, Choi Y. The Curious Case of Neural Text Degeneration. 2019.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eBen-Shabat I, Lindvall K, Salzer J. Exploring strategies for management of in-hospital stroke in Sweden: A qualitative study. PLoS ONE. 2024;19:e0313765.\u003c/span\u003e\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Offline AI-assisted qualitative analysis, Qualitative Research, Thematic Analysis","lastPublishedDoi":"10.21203/rs.3.rs-7122121/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7122121/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e\u003cp\u003eThe use of artificial intelligence (AI) in qualitative content analysis has increased substantially. However, most AI-assisted tools rely on proprietary cloud-based solutions, raising significant concerns regarding data protection, transparency, and reproducibility. Sensitive qualitative data such as interview transcripts require strict data sovereignty and compliance with legal and ethical standards, which existing cloud-based AI platforms cannot always guarantee. To address these challenges, this study aimed to develop and evaluate an accessible, fully offline AI tool tailored for qualitative research.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eWe developed QualiPilot, a stand-alone, privacy-compliant desktop application for AI-assisted qualitative analysis. The tool operates entirely offline, leveraging a locally deployed large language model (LLM) and requiring no installation, as it can be run directly from a USB flash drive. QualiPilot enables key steps of qualitative analysis, including initial text engagement, inductive category development, coding, and code summarization, with complete transparency of analytic parameters, system prompts, and procedural steps.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eQualiPilot performs all key analytic tasks on standard office computers without transmitting data to external servers, ensuring complete data protection and transparency. The tool produces relevant and well-structured outputs, with all steps, parameters, and prompts fully documented and exportable for reporting. The offline approach also simplifies compliance with data protection and ethical requirements. However, processing speed is slower compared to online or GPU-accelerated solutions, particularly for large datasets, and longer transcripts require manual segmentation due to model limitations. Despite these constraints, no instability or data loss is observed, and the tool proves practical for most typical qualitative projects.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e\u003cp\u003eQualiPilot provides a novel, privacy-compliant, and fully offline solution for AI-assisted qualitative content analysis. By enabling transparent, reproducible, and secure analytic workflows on standard hardware, the tool lowers barriers for researchers working with sensitive data. While not a replacement for manual analysis, QualiPilot offers valuable supplementary perspectives and supports rigorous, transparent qualitative research without dependence on commercial cloud services.\u003c/p\u003e","manuscriptTitle":"QualiPilot – an offline AI tool for qualitative content analysis","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-16 07:46:51","doi":"10.21203/rs.3.rs-7122121/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":"ac11d872-0328-4dca-b959-915b73c2b9af","owner":[],"postedDate":"July 16th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-07-17T15:38:10+00:00","versionOfRecord":[],"versionCreatedAt":"2025-07-16 07:46:51","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7122121","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7122121","identity":"rs-7122121","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}