AI Reporting Guidelines for Research: What Authors Need to Disclose and Report

Key Takeaways:

• Current reporting guidelines cover 2 separate duties: disclosing how AI was used as a research tool, and reporting the AI they evaluated as the object of study.
• The GAMER checklist supplies 9 items covering generative AI use in any medical study; PRISMA-trAIce covers AI used inside systematic reviews; SPIRIT-AI, CONSORT-AI, STARD-AI, MINIMAR, and CLAIM cover AI as the subject of research.
• The details editors ask for most often are the ones authors forget to capture during the study: model name, version, dates of use, prompts, settings, and verification steps.
• Journal rules remain inconsistent, so start tracking your AI use right from the time of designing the study protocol, rather than writing your AI declaration just by memory at the time of submission.

Introduction

AI now touches nearly every stage of a study: protocol drafting, literature searching and screening, data extraction, image analysis, statistical code, and manuscript language. Undisclosed or unverified AI use can produce fabricated citations, unreproducible methods, and, in serious cases, retraction. Journal editors and peer reviewers have therefore moved from asking whether AI was used to asking exactly what it did.

This article maps the guidelines that now govern that answer, and shows what to record while the study is running.

Why Does AI Disclosure Matter for Getting Published?

Disclosure has become a marker of methodological care rather than an administrative formality. Reviewers read the AI disclosure statement early, and a vague one invites scrutiny of everything that follows.

• Reproducibility: not reporting the model version, prompt, or temperature setting makes a methods section impossible to replicate.
• Integrity: AI-generated content that no human verified can carry false or exaggerated claims into the literature.
• Accountability: authors remain collectively responsible for every sentence, including AI-edited text.
• Screening efficiency: naming the affected sections speeds peer review instead of slowing it.

Are You Disclosing AI as a Tool or Reporting AI as a Subject?

Most confusion in this area comes from misunderstanding whether you’re using AI in your research or conducting research about AI.

AI as a research tool

• Covers using an AI tool for writing, outlining, code generation, data extraction, screening, and analysis support.
• GAMER applies here across all types of medical research, from study design through manuscript writing.
• For systematic reviews specifically, PRISMA-trAIce adds items for AI used inside the review workflow.
• Scope limits are worth noting: GAMER does not cover non-generative tools, general web search engines, or tools used solely for language translation.

AI as the intervention or test under evaluation

• Covers models that are themselves study variables: an AI triage system, a diagnostic algorithm, an imaging classifier.
• SPIRIT-AI, CONSORT-AI, STARD-AI, MINIMAR, CLAIM, TRIPOD+AI, and DECIDE-AI apply depending on design.
• PRISMA-AI sits in this category too, which is why it does not solve the tool-use problem; it was built around AI as a subject of research rather than as an instrument inside the review.
• AI can be used as a research tool AND can form the subject of research simultaneously. Supposing a researcher conducts a clinical trial of an AI-powered medical device, and then uses ChatGPT to draft some sections of the manuscript. That researcher needs to follow both CONSORT-AI and GAMER.

Major AI Reporting Guidelines

The table below summarizes scope and source for each guideline; the subsections add the items authors most often miss.

Guideline	Primary scope
SPIRIT-AI	Clinical trial protocols evaluating an intervention with an AI component
CONSORT-AI	Reports of clinical trials evaluating an intervention with an AI component
STARD-AI	Diagnostic accuracy studies using AI
MINIMAR	Medical AI research; data, methods, and study characteristics sections
CLAIM (2024 update)	AI in medical imaging studies
GAMER	Generative AI used as a tool in any type of medical research
PRISMA-trAIce	AI used as a methodological tool in systematic reviews

 

SPIRIT-AI: trial protocols involving AI

• Applies to the whole protocol, not a single section.
• Published simultaneously in BMJ, Nature Medicine, and Lancet Digital Health in 2020; Spanish and Portuguese translations are available.
• Specify before recruitment: model version, intended user, input and output handling, and where the tool sits in the care pathway.

CONSORT-AI: trial reports involving AI

• The reporting counterpart to SPIRIT-AI, and now read alongside the CONSORT 2025 statement.
• Report performance errors, human-AI interaction, and how poor-quality or missing input data were handled.
• State any change between the protocol version of the model and the version actually evaluated.

STARD-AI: diagnostic accuracy studies

• The newest of the design-specific extensions, published in Nature Medicine on September 15, 2025.
• Highest-risk omissions are the reference standard definition, data partitioning, and threshold selection.
• Report how cases with indeterminate or failed model outputs were counted.

MINIMAR: minimum information for medical AI

• Unusual among these guidelines in applying to specific report sections: data, statistical methods and analyses, and study characteristics.
• Emphasizes source population, demographics, model architecture, and evaluation; these are the fields that make bias auditable.
• Useful as a supplement when a design-specific extension says little about the training population.

CLAIM: AI in medical imaging

• The 2024 update revises the original 2020 checklist.
• Report acquisition parameters, the ground-truth annotation process, annotator expertise, and the de-identification method.
• Applies to the whole report and is widely used by radiology reviewers.

GAMER: generative AI use across all study types

• Developed through a scoping review, 2 Delphi rounds, and virtual consensus meetings.
• 51 experts from 26 countries took part, with 44 participating in the Delphi survey itself.
• Complements study-specific guidelines by covering every research phase rather than a single design.
• Detailed item by item in the next section.

PRISMA-trAIce: AI-assisted evidence synthesis

• Proposed as a discipline-agnostic extension to the PRISMA 2020 statement.
• Targets transparent reporting when AI acts as a methodological tool during evidence synthesis.
• Covered in depth later in this article.

The Nine Items of GAMER

Items 1 to 3: declaration, tool specification, and prompts

• Item 1: state whether any generative tool was used anywhere in the study or manuscript. If none was used, the remaining items do not apply, but an explicit negative statement is still recommended.
• Item 2: give the tool name, version or release date, and the dates or period of use. Say whether you used the front-end interface or the API, and report temperature, token length, language, or layer settings where available.
• Item 3: describe the prompting technique and provide the unedited responses, ideally as supplementary material submitted with the manuscript.

Items 4 to 6: custom models, role, and location in the manuscript

• Item 4: applies only if you developed or fine-tuned a tool. Report the name and version of the original base model. For widely recognized tools such as ChatGPT or Claude, this item is not applicable.
• Item 5: describe the functional role in every phase where the tool was used, such as language polishing, outlining, code generation, or analysis.
• Item 6: name the concrete sections or paragraphs affected. If the tool was used only for language editing across the whole text, listing sections is unnecessary.

The distinction between items 5 and 6 is worth stating plainly: item 5 is what the tool did, and item 6 is where in the manuscript it did it.

Items 7 to 9: verification, data privacy, and effect on conclusions

• Item 7: describe how generated content was proofread, checked, and revised. If no verification was performed, state the reason clearly.
• Item 8: describe anonymization, encryption, differential privacy, and compliance with local data protection rules. Remove sensitive patient information before any tool can access it.
• Item 9: state whether and how AI use may have influenced interpretation, accuracy, or conclusions, and confirm that all authors accept responsibility for the content.

Where GAMER-based disclosures are placed in the paper

Most GAMER items were suggested for the methods section, framed as a recommendation rather than a rule. The placement below works for most journals but always check the target journal first.

What you are reporting	Usual location
Tool name, version, dates, settings, role, and verification	Methods
Language editing only, with no effect on content	Acknowledgments or a short AI use statement
Prompts, transcripts, and unedited outputs	Supplementary material
Base model details for a fine-tuned tool	Methods, with technical detail in supplementary material
Effect of AI use on interpretation or conclusions	Discussion or limitations
Completed checklist	Uploaded with the submission files

What is likely to change next

• The GAMER team plans to approach editors of major journals about integrating the checklist into author guidelines, and a standing coordination group will review the statement annually.
• PRISMA-trAIce is pursuing formal endorsement through community consensus, so journal uptake will likely follow endorsement.
• Expect checklist upload at submission to become mandatory, as it already is for PRISMA and CONSORT.
• Monitor ICMJE recommendations, COPE position statements, publisher author guidelines, and funder mandates.

A Pre-Submission Checklist for Authors

• Identify the design-specific extension: SPIRIT-AI, CONSORT-AI, STARD-AI, CLAIM, MINIMAR, or TRIPOD+AI.
• For a systematic review that used AI in any review step, add PRISMA-trAIce alongside PRISMA 2020.
• Complete all 9 GAMER items if a generative AI tool was used in the research process at any stage.
• Archive prompts, outputs, model versions, settings, and dates while the study is running, not at write-up.
• Confirm that no identifiable patient data entered any tool, and record the anonymization method.
• Document who verified generated content and how discrepancies were corrected.
• Attach completed checklists as supplementary files; submitting the filled checklist is advised.
• Confirm that every author has reviewed the AI-assisted content and accepts responsibility for it.

Limitations and Unresolved Gaps

• Only 15% of the GAMER expert panel were women, and patient representatives were not included, which the developers flag as a priority for future updates.
• PRISMA-trAIce is a proposal awaiting formal consensus, so authors cannot yet cite it as an endorsed standard.
• Rapid changes and updates to AI models means version-specific reporting ages quickly, and deprecated versions may be unavailable for replication.
• No consensus yet exists on reporting agentic or multi-step AI workflows in which the model chains its own tasks.
• Guidance on retrieval-augmented systems and on privately hosted models remains thin.