Qualitative research serves a foundational role in health sciences by providing deep, contextualized insights into human behaviors, perceptions, and lived experiences [1-3]. Unlike quantitative approaches, qualitative studies use non-numerical methods, enabling detailed exploration and generating novel perspectives within health care [1,4]. Qualitative research encompasses diverse methodologies, predominantly involving interviews, focus groups, participant observation, and analysis of documents or artifacts, supported by analytic techniques, such as coding, thematic analysis, and content analysis [5,6].

Ensuring rigor and transparency in qualitative research requires standardized reporting guidelines. Since its publication in 2007, the COREQ (Consolidated Criteria for Reporting Qualitative Research) checklist has become widely accepted, with endorsements from organizations such as the Enhancing the Quality and Transparency of Health Research (EQUATOR) Network, and has been cited more than 30,000 times [7]. COREQ supports methodological transparency and clarity and, consequently, the interpretability and trustworthiness of qualitative findings.

However, qualitative research methodologies are evolving rapidly due to the accelerated integration of large language models (LLMs), a type of artificial intelligence (AI). These technologies introduce both unprecedented opportunities and novel methodological challenges beyond the scope of the original COREQ guidelines. LLMs, trained on extensive text corpora, generate coherent, contextually relevant text in response to user inputs (prompts) [8]. Their capacity to analyze, interpret, and generate humanlike language underpins their increasing use across qualitative research and practical applications.

The integration of LLMs into qualitative research workflows is expanding rapidly. Researchers use these models at multiple stages, including formulating research questions; designing interview protocols; supporting data processing tasks, such as transcription and translation; aiding analytic processes, such as coding and theme identification; drafting manuscript sections; and facilitating interactive dialogue with qualitative datasets [9-19]. This adoption is largely motivated by the perceived benefits offered by LLMs [20]. Notably, LLMs can enhance the speed and efficiency of data analysis by rapidly processing extensive textual datasets, enabling researchers to undertake more expansive and expedited investigations. Beyond efficiency, LLMs provide novel analytic perspectives by identifying latent semantic patterns, proposing new thematic frameworks, and distilling complex narrative content. These capabilities complement human interpretive efforts and have potential implications for reducing interpretive biases. Consequently, leading qualitative analysis software packages, including ATLAS.ti and MAXQDA, now integrate LLM functionalities [21,22].

Despite these advantages, the integration of LLMs into qualitative research introduces several methodological risks and necessitates rigorous reporting standards. Unlike conventional software tools that often rely on deterministic algorithms, LLMs function probabilistically. While adept at identifying and reproducing language patterns from training data, their semantic comprehension, particularly regarding nuanced contexts and subtexts, remains uncertain and under active scientific investigation [23]. Furthermore, because LLMs are trained on large-scale datasets that reflect statistical regularities within human language, they are susceptible to reinforcing existing biases, perpetuating dominant cultural narratives, and systematically underrepresenting marginalized perspectives [24,25]. The internal mechanisms guiding their outputs are often opaque (“black box” phenomenon), which poses significant challenges for interpretability and explainability. In addition, LLMs are prone to producing outputs that are syntactically plausible yet factually inaccurate, a behavior known as “hallucination” [26]. In the context of qualitative research, such inaccuracies may be especially difficult to detect due to interpretive depth and narrative complexity. These factors collectively introduce critical challenges to reproducibility, validity, and overall trustworthiness of qualitative research facilitated by LLMs.

Reflecting these concerns, the qualitative research community is actively debating the appropriate role and regulation of LLMs. Schroeder et al [27], after interviewing 20 qualitative researchers, highlighted an “urgent lack of norms” governing model use and advocated for tool-supported transparency. Zhang et al [28] further argued that enhancing transparency and deepening researchers’ understanding of LLM capabilities can improve their effective use. Similarly, in a workshop involving 40 scholars, Quéré et al [29] observed that LLMs have already been integrated into many researchers’ workflows; however, they emphasized the need for defined standards to assess validity and rigor that, at present, remain undefined.

The existing COREQ checklist offers insufficient guidance for navigating these complexities. Its only applicable item (item 27: “What software, if applicable, was used to manage the data?”) inadequately captures essential methodological details needed to assess the implications of LLM use. Merely naming software or models used is insufficient; understanding precise use contexts, limitations, and verification strategies is crucial for evaluating research quality.

Existing reporting guidelines for AI, such as TRIPOD-LLM (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis–Large Language Model), CONSORT-AI (Consolidated Standards of Reporting Trials-Artificial Intelligence), and SPIRIT-AI (Standard Protocol Items: Recommendations for Interventional Trials-Artificial Intelligence), primarily address areas that are not specific to qualitative research. Instead, these guidelines focus on topics such as the development or evaluation of LLMs, randomized controlled trials involving AI, and clinical trial protocols evaluating interventions with AI, respectively [30-32].

Recognizing this critical methodological gap, we are developing COREQ+LLM, a rigorous extension of the established COREQ checklist. COREQ+LLM will introduce specific reporting items essential for transparently describing and critically evaluating qualitative research incorporating LLMs. This extension aims to capture detailed methodological information, enabling a thorough appraisal of rigor, trustworthiness, and reproducibility, thereby maintaining the foundational integrity of qualitative research in the era of AI.

The overall objective of this study is to extend the existing COREQ checklist to support transparent and responsible reporting of LLMs in qualitative research. To achieve this, we aim to (1) identify and categorize current applications of LLMs used as qualitative research tools (eg, conducting interviews, analyzing focus group discussions, and extracting themes from textual data; refer to Inclusion and Exclusion criteria section for more details), (2) assess how LLM use as a qualitative research tool is reported in peer-reviewed empirical studies, and (3) develop and refine reporting items for COREQ+LLM through a structured consensus process among international experts.

Compared with earlier natural language processing (NLP) approaches, such as statistical n-gram models or task-specific neural networks, LLMs function as general-purpose problem-solvers [33]. Their breadth of capability arises from massively scaling model parameters, training data, and compute used for training [34]. While traditional NLP pipelines were engineered for narrowly defined tasks, such as feature extraction, LLMs can generalize across an extensive array of linguistic tasks, which they were never explicitly trained to perform. Therefore, focusing the COREQ extension on LLMs is warranted on both technical and pragmatic grounds. From a technical perspective, LLMs possess generative capacities that create entirely new epistemic objects (eg, full paragraphs, code blocks, or analytic summaries) rather than merely classifying or extracting text. These capabilities introduce LLM-specific risks that are intrinsic to the model architecture, notably hallucinations and nondeterministic outputs [35]. From a pragmatic standpoint, expanding our checklist to encompass other AI modalities, such as computer vision, reinforcement learning, symbolic AI, or rule-based NLP, would render the resulting framework overly complex and risk impeding consensus among researchers. Given that qualitative research fundamentally engages with linguistic artifacts, such as interviews and documents, maintaining a language-first focus ensures that our guidance remains aligned with the empirical substrate of qualitative inquiry while avoiding an unwieldy, modality-agnostic approach.

At the same time, the integration of LLMs into qualitative research represents part of a broader trend toward the adoption of AI in qualitative research. By focusing on LLMs, COREQ+LLM provides a timely and targeted response to the most prominent and methodologically disruptive form of AI currently impacting qualitative inquiry. This extension can serve as a foundational step toward broader guidance on AI-supported qualitative research while ensuring actionable and consensus-driven reporting standards for the most immediate use cases. For these reasons, we propose the guideline extension to be titled “COREQ+LLM” rather than “COREQ+AI.”

The development of the COREQ+LLM extension adheres to internationally recognized methodologies for creating health research reporting guidelines, primarily drawing upon the guideline development approach provided by the EQUATOR Network [36]. This approach aligns with precedents set by other AI-focused reporting guideline extensions, such as TRIPOD-LLM, CONSORT-AI, and SPIRIT-AI [30-32]. Recognizing the complexity and rapid evolution of LLM technologies, we use a rigorous, multiphase methodology, comprising a scoping review (phase 1) and a Delphi consensus process (phase 2). Given the rapid evolution of the use of LLMs in qualitative health research and the lack of standardization, a Delphi method allows for structured expert deliberation while integrating emergent, interdisciplinary knowledge [37]. Table 1 summarizes the research process.

This study will be overseen by a 6-member steering committee comprising AI researchers, clinicians, ethicists, experts in qualitative research methodology, and representatives for good scientific practice from Witten/Herdecke University. The protocol will be registered with the EQUATOR Network as a reporting guideline under development. The scoping review protocol was registered with the Open Science Framework on June 6, 2025 [38], and adheres to the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) guidelines [39].

The steering committee will oversee the development of the COREQ+LLM explanation and elaboration document, with draft versions circulated to consensus meeting participants for review and feedback. To enhance clarity and usability, pilot testing will be conducted with 3 to 5 researchers, including PhD students, who are currently conducting qualitative research involving LLMs as a research tool. Feedback will be used to identify any items that are unclear or difficult to interpret. Participants will apply the draft checklist to a relevant study or their own work and provide structured feedback on item clarity, relevance, and ease of use. On the basis of this input, the steering committee may revise the checklist (adding, modifying, or removing items) as necessary to ensure the final version is both comprehensive and practical.

We plan to disseminate COREQ+LLM widely by publishing it in a peer-reviewed journal, presenting it at major academic and clinical conferences, and making it freely accessible to the global research community through the EQUATOR Network website [45].

The ethics committee of Witten/Herdecke University raised no objection regarding ethical concerns (S-165/2025). Each participant will be asked to provide informed consent prior to their participation in the Delphi study. No financial incentives will be offered to participants.

The integration of LLMs into qualitative research signifies a substantial shift in methodologies for data collection, analysis, and interpretation. Without specialized reporting guidance, there is a risk to the quality, transparency, and trustworthiness of research involving LLMs. We anticipate the scoping review will reveal variability and gaps in reporting practices concerning the use of LLMs, reinforcing the need for structured guidelines. Given these challenges, we expect great stakeholder interest in the Delphi process, as the absence of clear reporting standards constitutes a considerable burden. COREQ+LLM aims to bridge this critical gap by offering a comprehensive, consensus-driven framework for transparently reporting the use of LLMs in qualitative research.

The COREQ+LLM guideline is designed to support diverse stakeholders. Researchers will gain structured guidance for transparently documenting LLM use; peer reviewers and journal editors will have a robust tool to evaluate the completeness and methodological rigor of submissions; clinicians, policymakers, patients, and other researchers will benefit from enhanced clarity and confidence when interpreting LLM-assisted qualitative findings. Ultimately, COREQ+LLM seeks to improve the credibility, validity, and practical relevance of qualitative research using advanced LLM-supported methodologies.

To the best of our knowledge and based on a review of the “reporting guidelines under development for other study designs” section of the EQUATOR Network, there are no reporting guidelines that specifically address the use of LLMs in qualitative research [46]. AI-focused extensions of reporting guidelines exist (eg, CONSORT-AI, TRIPOD-LLM, and SPIRIT-AI); however, they do not address the specific methodological implications of using LLMs in qualitative inquiry. Existing guidelines, such as COREQ, contain only minimal references to software use and do not account for the probabilistic, generative, and opaque characteristics of LLMs. COREQ+LLM addresses this gap by focusing on the use of LLMs as qualitative research tools rather than their evaluation or perception.

A major strength of this protocol is its adherence to EQUATOR and DELPHISTAR guidance, ensuring methodological rigor in both the scoping review (registered with the Open Science Framework [38]) and the Delphi consensus process. One potential limitation is that expert participation in Delphi studies can vary; therefore, strategies to mitigate low response rates have been included. Furthermore, the scoping review will be restricted to English-language publications, and although the steering committee will be purposefully composed to reflect diverse geographic regions and disciplines, it may not encompass all perspectives. These factors may influence the generalizability of the final checklist.

Upon completion, COREQ+LLM may serve as a foundation for broader extensions addressing other AI modalities in qualitative research (eg, computer vision). It may also stimulate further methodological research into the integration of LLMs and hybrid human-AI analytic workflows in qualitative research studies.

The results from this study, including the final COREQ+LLM checklist and explanation and elaboration document, will be published in a peer-reviewed journal and submitted to the EQUATOR Network for public use (refer to phase 4: Dissemination section). Moreover, we will plan to present findings at relevant academic and clinical conferences to promote awareness and uptake across the qualitative health research community.

COREQ+LLM responds to the growing use of LLMs as a qualitative research tool by providing structured, consensus-based guidance for transparent reporting. This protocol outlines a rigorous, multiphase development process that combines systematic evidence synthesis with expert consensus through a Delphi process. The resulting guideline aims to enhance trust, reproducibility, and methodological transparency in LLM-assisted qualitative research.