The GAMER Statement: How to Report Generative AI Use in Medical Research

There are now two quite different problems that journals are trying to address when it comes to generative AI. The first, which has received most of the attention, is the disclosure problem: did you use a large language model to help write or revise your manuscript, and have you said so? The second problem is less widely discussed but, for research validity, arguably more important. Did you use a generative AI tool as part of the research itself, and if so, did you report that use with enough methodological transparency that someone else could reproduce your work?

Those are different questions. A paper can include a clean AI disclosure statement in its acknowledgements and still provide essentially no reproducible detail about how ChatGPT was actually prompted, which version was used, whether outputs were verified, or how the tool's role related to the paper's conclusions. That methodological gap is what the GAMER Statement was designed to close.

The GAMER Statement (Reporting guideline for the use of Generative Artificial intelligence tools in MEdical Research) was published in BMJ Evidence-Based Medicine in December 2025, in Volume 30, Issue 6, pages 390 to 400. It was developed by 51 experts from 26 countries using a Delphi consensus process and registered with the EQUATOR Network, the body that houses reporting standards like CONSORT, PRISMA, and STROBE. The result is a 9-item checklist that applies whenever a generative AI tool plays any substantive role in any phase of a medical study, from data synthesis and content extraction to clinical text classification, qualitative coding, or systematic search assistance.

Why a Separate Reporting Standard Was Needed

The problem that prompted GAMER's development was not theoretical. Systematic reviews of published papers using ChatGPT and similar tools consistently found that the majority of studies failed to report key methodological details: the version of the model, the date queries were run, the exact prompts used, whether questions were asked in a single session or across multiple independent chats, and how or whether outputs were verified before being incorporated into the analysis. In some analyses, fewer than 40 percent of ChatGPT studies reported the core elements of their query strategy.

This creates a real reproducibility problem. A study that asks GPT-4o whether a set of clinical trials meets eligibility criteria for a systematic review produces different results depending on which version of the model is used, what temperature setting it runs at, how the prompt is worded, and whether context from prior turns in the conversation influences later answers. If the paper does not report these details, no subsequent researcher can replicate the procedure. Worse, readers have no way to assess how sensitive the findings might be to those choices.

The problem is compounded by the speed at which generative AI models are updated. A study conducted with GPT-3.5-turbo in early 2024 cannot be directly compared to a study conducted with GPT-4o in late 2025. The models differ in their training data, parameter counts, fine-tuning approaches, and refusal behaviors. When the version is not reported, even the direction of any difference is unknowable.

Existing journal AI policies addressed the authorship and disclosure questions reasonably well. The ICMJE, COPE, and the major publisher policies converged on the same core principle: AI tools cannot be authors, and substantive AI use must be disclosed. But those policies were not designed to specify what methodological information a researcher must provide when a large language model serves as a data processing tool, a content synthesizer, or a decision-support aid within the research itself. The GAMER Statement fills that gap.

The 9 GAMER Checklist Items, Explained

The checklist is structured so that each item corresponds to a specific piece of methodological information that either editors, peer reviewers, or future readers would need to understand or replicate the study. Some items will feel straightforward; others will require more deliberate documentation from the outset of a project.

Taken together, these nine items describe a methodological transparency standard that is broadly analogous to what CONSORT expects from randomized trials or STROBE from observational studies. The principle is the same: readers and reviewers need enough information to evaluate the study's validity independently.

What Studies Are Covered

The GAMER Statement is designed to apply whenever a generative AI tool plays a substantive role in any phase of medical research. The developers defined the scope deliberately broadly, covering the full range of ways these tools are now being used: as literature screening aids in systematic reviews, as text classifiers for free-form patient data, as draft generators for clinical summaries, as prompt-driven diagnostic support tools under evaluation, and as research assistants for qualitative coding, thematic analysis, and content extraction.

What the guideline does not address, by design, is incidental AI use in manuscript writing that has no bearing on the study's research methodology. If a researcher used Claude or ChatGPT to improve the English in the discussion section after the analysis was complete, that belongs in an AI disclosure statement but does not require a full GAMER report. The boundary between the two can sometimes feel unclear, but a useful test is whether the AI tool's outputs influenced any data, interpretation, or claim that appears in the results section. If yes, GAMER applies.

The checklist also applies, with different emphasis, to studies that evaluate generative AI tools rather than simply using them. A paper testing whether GPT-4o can correctly triage emergency department referrals, or whether a large language model can accurately summarize clinical trial results, is itself a study of AI performance. Those papers require full GAMER reporting for obvious reasons, because the methodological choices (which prompts, which model version, which verification procedures) are the study's independent variables.

Where GAMER Fits Among Other AI Reporting Guidelines

The GAMER Statement is not the only reporting tool in this space, and understanding how it differs from related guidelines matters for choosing which one applies to your paper. The landscape of AI reporting standards in medicine is consolidating relatively quickly, but the field is still new enough that overlap between guidelines creates some real confusion at submission time.

The CHART Statement (Chatbot Assessment Reporting Tool), published in 2025 in both BMC Medicine and JAMA Network Open, was developed specifically for studies that evaluate chatbot performance in providing health advice, clinical summaries, or patient-facing information. CHART has 12 items and 39 subitems and goes into considerable depth on how queries were structured, how accuracy was assessed, and how the chatbot's outputs were compared against reference standards. Where CHART applies to a study, it should be used instead of, or alongside, GAMER depending on overlap.

TRIPOD+AI, published in the BMJ in April 2024, is a reporting standard for clinical prediction models that use machine learning or AI, including large language models used as components of a predictive pipeline. TRIPOD+AI covers model development, validation, and deployment reporting in ways that go beyond what GAMER addresses. If your study is primarily about building or validating a clinical AI model, TRIPOD+AI is the primary guideline, though GAMER's items on prompting and verification remain relevant for any generative components.

CONSORT-AI and SPIRIT-AI extend the clinical trial reporting standard to studies where AI is the intervention under evaluation. These have been endorsed by journals including The Lancet, Nature Medicine, and BMJ since their publication in 2020, and they cover randomized trial design and protocol reporting specifically.

GAMER sits in the gap that these more specialized guidelines do not cover: the broad middle category of medical research studies where generative AI tools are used as research instruments without the paper being primarily a trial, prediction model study, or chatbot evaluation. As that category now accounts for a rapidly growing share of published medical AI research, GAMER's scope makes it the default starting point for most authors in this position.

The Prompt Documentation Problem

Of all nine GAMER items, the one that researchers are least prepared for is number three: prompting techniques. This is also the item that, when missing, most directly undermines reproducibility.

A large language model is not a fixed algorithm. The same model, given slightly different instructions, will produce different outputs, sometimes very different ones. A prompt that asks a model to "identify studies meeting inclusion criteria based on the following abstract" will behave differently from a prompt that asks it to "read the following abstract and decide Yes or No whether this study meets the following inclusion criteria, then explain your reasoning." The second formulation uses chain-of-thought reasoning, which tends to improve accuracy on structured classification tasks. The first does not. If neither the original researchers nor subsequent readers know which approach was used, the study's classification results cannot be meaningfully interpreted or reproduced.

This matters more as studies using generative AI methods are being published in high-impact venues. When a paper in The Lancet or JAMA reports that a large language model achieved 88 percent accuracy in some clinically meaningful task, readers need to know exactly what prompt produced that result. A different prompt on the same model might achieve 65 percent or 94 percent on the same task. The headline number is meaningless without the prompt.

Practically, this means researchers should be maintaining a prompt log from the start of a project. Every significant prompt used in any analytical step should be saved, versioned, and stored as part of the study's documentation. The exact text should be reported in the methods or as a supplementary file. If a prompt was refined through an iterative development process, describing that process is part of the GAMER report. Some journals are now asking for prompts to be deposited in a data repository alongside code and raw data. Researchers who built their study on informal prompt experimentation without documentation will find this requirement difficult to meet after the fact.

Data Privacy and Patient Information

Item 8 in the GAMER checklist, data privacy, is the one most likely to create legal as well as editorial problems for researchers who have not thought it through in advance.

Most commercial large language model APIs process inputs on the vendor's servers. When a researcher submits clinical notes, patient-reported outcomes, or hospital records to a consumer-facing AI tool, those data leave the institution's controlled environment. Depending on the tool's terms of service and data retention policies, inputs may be used to train future model versions. Depending on the jurisdiction and the nature of the data, submitting patient information to a third-party service without explicit institutional review board approval and patient consent may constitute a breach of data protection requirements, including HIPAA in the United States and GDPR in Europe.

Some researchers work around this by using enterprise-tier API access with data processing agreements that exclude training use and include contractual confidentiality protections. Others use locally-deployed open-weight models that process all data within institutional infrastructure. Others de-identify data before submission to external services, though the sufficiency of that de-identification depends on both the de-identification method and the sensitivity of the specific data.

GAMER's data privacy item requires authors to describe which of these approaches they used. For editors, this information is important because papers that used patient data in a third-party AI tool without appropriate governance may need to be sent back for ethics review regardless of the study's scientific merits. For readers, it signals whether the research was conducted with the kind of data stewardship that makes clinical AI research credible.

Content Verification: What Human Review Actually Means

Item 7, content verification, is where the GAMER checklist goes beyond what disclosure policies address. Stating that a human reviewed AI outputs before they entered the analysis is not sufficient on its own. The checklist requires a description of how that review was conducted, what proportion of outputs were checked, and what happened when the review found errors.

In practice, verification standards vary enormously across published AI-assisted studies. Some papers report that all AI outputs were independently reviewed by two researchers with inter-rater reliability calculated against the AI labels. Others describe a spot-check of a randomly selected 10 percent sample. Others acknowledge that outputs were "reviewed by the authors" without specifying the review procedure. GAMER makes that variability visible and requires it to be reported rather than elided.

For systematic reviews using AI to screen titles and abstracts, the verification question connects directly to PRISMA reporting: if a model made preliminary eligibility decisions, what were the false negative rates, and how do those rates affect the completeness of the final included evidence base? For clinical text classification studies, the verification question connects to standard psychometric reporting: what is the agreement between human and AI coders, and at what threshold were discrepant cases resolved?

Researchers who treated generative AI as a fully trusted data processor rather than a probabilistic tool that requires checking will struggle to satisfy this item. The honest solution, at this stage of the technology, is to build verification into the research protocol before data collection begins rather than adding a description of whatever informal review happened to occur.

What Journal Editors Are Starting to Ask For

The GAMER Statement was developed under EQUATOR methodology, which means it was designed from the beginning to be implementable as a submission requirement, not merely an aspirational guideline. The testing phase began in early 2025 and continues. As of mid-2026, the checklist is not yet universally mandated at major medical journals in the way CONSORT or PRISMA are, but editors at journals publishing AI-related research are increasingly aware of it and referencing its items in their review criteria.

BMJ Evidence-Based Medicine, which published the GAMER Statement, is an obvious venue where editors will apply the checklist. Journals publishing systematic reviews and meta-analyses are also natural early adopters, since AI-assisted screening is now common in that study type and the reproducibility stakes are high. Specialty journals in fields where AI clinical applications are being actively evaluated, including radiology, pathology, ophthalmology, and emergency medicine, are paying close attention.

The practical consequence for authors is that a paper using generative AI tools in its research process may now receive reviewer comments asking for specific GAMER items even at journals that have not formally adopted the checklist. That kind of informal diffusion, where a guideline's criteria become part of reviewer expectations before journal policy catches up, is exactly how CONSORT and PRISMA spread in the early 2000s. Being ahead of that adoption curve means fewer revision cycles and stronger credibility in peer review.

Authors targeting high-impact journals should also be aware that the absence of prompt documentation is increasingly treated as a methodological deficiency, not merely a reporting lapse. A paper claiming that a large language model outperformed clinicians at some task, without specifying the exact prompts used, presents an unverifiable finding. Reviewers at journals like The Lancet Digital Health, JAMA Network Open, and npj Digital Medicine have been explicit about this in published editorial correspondence since late 2025.

Where to Find the GAMER Checklist

The GAMER Statement can be found on the EQUATOR Network website under AI and machine learning study designs, where it is registered alongside CHART, TRIPOD+AI, and other related standards. The primary publication is in BMJ Evidence-Based Medicine (Volume 30, Issue 6, December 2025, pages 390 to 400), which is the authoritative reference for citation purposes. The development protocol was published separately in JMIR Research Protocols in 2025.

The EQUATOR listing includes the checklist itself as a downloadable document, which authors can use as a self-assessment tool before submission. Working through the checklist at the point when the methods section is being finalized, rather than at the very end of the writing process, will reveal any documentation gaps while there is still time to fill them from project records. Waiting until reviewer comments arrive is a much harder position from which to reconstruct prompt logs, model version information, and verification procedures.

A Practical Preparation Checklist for Authors

The researchers who will find GAMER compliance easiest are those who designed their AI data collection procedures the same way they design any other measurement instrument: prospectively, with documentation built into the protocol from the start. For studies already in progress or approaching submission, the following steps can recover most of the required information if records were kept.

A paper that can answer all nine of these questions clearly and specifically is a paper that a reviewer can evaluate properly. That is the standard the GAMER Statement sets, and it is a reasonable one given the reproducibility stakes involved in publishing findings that depend, in part, on how a probabilistic language model was prompted and used.

The Broader Shift This Represents

The GAMER Statement is part of a broader normalization of generative AI as a research instrument. A few years ago, using ChatGPT in a medical study was unusual enough that editors treated it as a novelty requiring explanation. Now it is common enough that journals need a structured way to evaluate whether the use was methodologically sound.

The analogy to other methodological standards is instructive. Statistical reporting in medicine was similarly informal for decades. Studies reported what they found without specifying their analysis plans in advance, without reporting effect sizes alongside p-values, without describing how missing data were handled. The SAMPL guidelines and subsequent pressure from journals on statistical transparency improved that situation considerably, though compliance is still imperfect. GAMER is the beginning of the same process for AI methods: establishing a floor of transparency that can be checked, enforced, and improved over time.

The implication for researchers is forward-looking. The GAMER checklist is, at present, a useful tool for planning good documentation habits. Within the next two or three years, as journals formalize their requirements for AI-assisted research reporting, it is likely to become a submission requirement at the venues most likely to publish this type of work. Building documentation practices now, before the formal mandates arrive, means less retrofitting under deadline pressure and more time spent on the science.

The researchers who are doing this well already treat prompt development as part of their methods section, not an afterthought. They save model version information alongside their dataset version information. They run their AI-assisted classification steps with the same inter-rater verification they would apply to any other subjective coding task. That is not more work than their peers are doing. It is the same work, done with documentation from the start rather than scrambled together at the revision stage.

Further Reading