An Interactive Qualitative-AI Analysis Protocol [working draft] Dr Ben Laws Dept of Psychiatry University of Cambridge Status: Early Development Version Intended Use: Interactive, Open-access, methodological resource [December, 2025] Abstract Large language models are increasingly being used to support qualitative data analysis, yet there is limited guidance on how such tools can be integrated in ways that are methodologically coherent, ethically defensible, and practically useful for researchers. This report introduces the Interactive Qualitative-AI Analysis Protocol, a decision-guided framework designed to support the responsible use of AI in qualitative research and to increase quality of outputs. The protocol leads researchers through a structured set of choices concerning analytic purpose, methodological stance, level of AI involvement, workflow design, trustworthiness, and transparency. It also provides practical resources for prompt design, multi-step analytic strategies, transcript preparation, and the detection of errors and hallucinations. Rather than prescribing a single analytic approach, the protocol functions as a flexible scaffold that accommodates diverse qualitative methodologies while foregrounding reflexivity, auditability, and human interpretive authority. The protocol is presented in its early development form and will be released open access, with the intention of ongoing refinement informed by user feedback and methodological debate. Recent advances in large language models (LLMs) have generated considerable interest in their potential to support qualitative data analysis, including summarisation, coding, pattern identification, and analytic memo writing (Gilardi et al., 2023; Nelson, 2024). Empirical studies suggest that such tools may substantially increase efficiency and scalability in qualitative research, particularly for large or complex datasets. However, despite this promise, there are currently few established methodological guides for how AI should be integrated into qualitative analysis in a way that is rigorous, transparent, and ethically defensible. A growing body of methodological work highlights that the reliability and usefulness of AI-assisted qualitative analysis are highly sensitive to prompt design, task decomposition, and analytic framing (Zhang et al., 2024; Deng et al., 2023). Poorly specified prompts can lead to hallucinations, over- generalisation, or the imposition of analytic structures that are misaligned with the researcher’s epistemological stance. At the same time, Zhang et al. (2024) show that for some researchers, particularly those new to AI-assisted methods, the time and cognitive effort required to design and refine prompts can offset anticipated efficiency gains, creating a tension between methodological care and practical feasibility. Figure 1. An overview of the 13 steps in the interactive protocol Alongside these technical concerns, qualitative researchers frequently raise ethical and epistemic questions about trust, transparency, interpretive authority, and participant misrepresentation, especially in sensitive domains such as mental health, clinical research, social care, and justice contexts (Bender et al., 2021; Birhane et al., 2023). These concerns are not merely technical but relate to core qualitative principles such as reflexivity, credibility, and accountability. The Interactive Qualitative-AI Analysis Protocol has been developed in response to these challenges. Rather than proposing a single ‘correct’ way to use AI, the QAP provides an interactive, decision- guided framework that helps researchers rapidly clarify: • their analytic purpose and methodological stance, • the appropriate level of AI involvement, • the safeguards required to maintain trustworthiness, • and the kinds of prompts and workflows most suited to their project. Crucially, the protocol is designed so that researchers can arrive at a coherent and defensible prompting strategy within minutes, thereby reducing the burden of prompt engineering while retaining methodological control. Built-in safeguards explicitly address issues of hallucination detection, auditability, and transparency. The term protocol is used deliberately but loosely. Qualitative research encompasses a wide range of epistemological traditions and analytic practices. A rigid, linear protocol would risk narrowing this diversity. Instead, the QAP functions as a structured scaffold that supports choice, reflexivity, and methodological alignment. The intention is not to standardise qualitative analysis, but to support responsible and informed engagement with AI tools. The protocol is currently in early development and will be released open access. It is intended to undergo very phases of iterative refinement informed by user feedback (from qualitative experts and prompt engineering specialists), rigorous empirical testing, and ongoing methodological debates. Overview of the Protocol Structure Sections 1–7 form the core analytic pathway and are intended to be completed sequentially. Together, they guide users through key methodological and ethical decisions before substantive AI-assisted analysis begins. Sections 8–13 function as supporting resources, offering prompt templates, workflow examples, and practical guidance for transcript preparation and error checking. The protocol is interactive: as users select options within each section, customised text is generated, allowing them to build a project-specific analytic and prompting strategy that can be directly reused. Figure 2. Showing the selectable sub-prompts that uses can choose to help orientate their analysis. The following table below gives a more detailed overview of the 13 sections of the protocol. Section Focus Rationale Benefit to Researchers 1. Project Purpose Exploratory, descriptive, interpretive, evaluative aims AI behaves differently depending on analytic intent; misalignment risks over- or under-interpretation Ensures AI use is appropriate to analytic goals and stage of research 2. Safeguarding & Data Governance Privacy, ethics, institutional compliance Ethical and legal concerns are major barriers to AI adoption Encourages lawful, ethical, and auditable AI use from the outset 3. Level of AI Involvement Light-touch to automation-oriented use AI can play multiple roles; conflating them obscures methodological risks Supports proportional, intentional delegation of analytic labour 4. Methodological Stance TA, IPA, grounded theory, framework analysis AI use must respect epistemological assumptions Strengthens methodological coherence and credibility 5. Workflow Type Descriptive, reflexive, minimal- AI workflows AI can intervene at different analytic moments Helps integrate AI without displacing human judgement 6. Validation & Trustworthiness Hallucination audits, triangulation, repeatability Traditional qualitative criteria need reinterpretation for AI contexts Enhances credibility, dependability, and analytic transparency Section Focus Rationale Benefit to Researchers 7. Output & Transparency Reporting AI use Journals and funders increasingly expect disclosure Supports defensible and transparent methods reporting 8. Prompt Models Cheat-Sheet Reusable micro- templates Prompt design is a major bottleneck for researchers Reduces time spent designing prompts while improving quality 9. Prompting Strategies No-shot, few-shot, chain-of-reasoning Different strategies entail different risks and benefits Enables informed choice between creativity, consistency, and auditability 10. Decomposition & Multi-Step Prompting Layered analytic pipelines Complex tasks increase hallucination risk Improves interpretive safety and mirrors qualitative logic 11. Self-Critique & Adversarial Review Reflexive and critical prompts AI outputs require systematic challenge Supports negative case analysis and bias detection 12. Combined Workflows End-to-end prompting strategies Isolated prompts are less effective than coherent workflows Offers ready-to-use analytic pathways for different research contexts 13. How to Prepare Transcripts Input quality and verification Many AI errors arise from poor inputs or traceability Improves accuracy, quote verification, and error detection At the end of the protocol, an automatically generated free-text box collates the options selected across the 13 sections. The purpose of this feature is to support researchers in translating methodological decisions into a project-specific prompting strategy. As users interact with the protocol, the text box incrementally assembles structured guidance and example language that can be adapted into bespoke prompts for their own qualitative datasets. Rather than supplying fixed prompts, this design encourages researchers to actively author and refine prompts that reflect their analytic aims, methodological stance, and safeguarding requirements, thereby maintaining human interpretive control while reducing the cognitive burden of prompt construction. Figure 3. A free-text box which accumulates user selections (example selections given). This text will help users to form their own unique study prompts. Conclusion This report has presented the Interactive Qualitative-AI Analysis Protocol as an early-stage methodological framework for supporting the responsible use of AI in qualitative research. The protocol is not intended as a definitive or closed system, but as a structured starting point that brings together emerging insights from qualitative methodology, prompt design, and concerns about trust, transparency, and analytic integrity. Its central contribution lies in offering researchers a decision-guided scaffold that helps translate methodological intent into coherent and auditable AI-assisted analytic practices. As an early version, the protocol necessarily has limitations. Further refinement is required to test its usability, clarity, and robustness across a wide range of qualitative methodologies, disciplines, and dataset types. In particular, systematic “stress testing” of the protocol on diverse qualitative datasets will be essential to examine how well it supports different analytic goals, levels of sensitivity, and research contexts, as well as to identify points where additional guidance, constraints, or flexibility may be needed. The protocol is therefore conceived as an iterative and evolving resource. Future development will focus on empirical evaluation, user feedback, and methodological critique, with the aim of refining both the conceptual structure of the protocol and its practical implementation. Through successive iterations, the longer-term ambition is to contribute to the development of shared standards and good practice for qualitative-AI analysis, while preserving the diversity and reflexivity that are central to qualitative research. Acknowledgements The development of the Interactive Qualitative-AI Analysis Protocol has been supported by the Accelerate Programme for Scientific Discovery, whose funding has enabled the initial design and implementation of this early-stage methodological resource. References Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of the ACM Conference on Fairness, Accountability, and Transparency. Birhane, A., et al. (2023). The values embedded in machine learning research. AI & Society, 38, 1287– 1300. Deng, J., et al. (2023). Large language models for qualitative research: Opportunities and challenges. arXiv preprint arXiv:2309.xxxxx. Gilardi, F., Gessler, T., Kubli, M., & Müller, S. (2023). ChatGPT outperforms crowd workers for text- annotation tasks. Proceedings of the National Academy of Sciences, 120(30). Nelson, L. K. (2024). Computational grounded theory: A methodological framework. Sociological Methods & Research. Zhang, Y., et al. (2024). Prompting trade-offs in AI-assisted qualitative analysis. Journal of Mixed Methods Research, advance online publication.