Gastroenterology's AI Policy: AGA and Elsevier Rules for GI Authors

Endoscopy and computer-aided detection

Gastroenterology publishes significant research on AI-assisted endoscopy. Computer-aided polyp detection (CADe), computer-aided diagnosis (CADx), AI for Barrett's esophagus surveillance, capsule endoscopy reading algorithms, these are among the journal's most-cited article types.

If your paper is about an AI endoscopy system, the AI is your research subject. It belongs in your standard Methods section with details on architecture, training data, validation design, and performance metrics.

The manuscript preparation AI disclosure is entirely separate. If you built a novel AI system for colorectal polyp characterization and also used Claude to edit your Discussion, you need two distinct descriptions that don't bleed into each other.

Example for an endoscopy AI paper:

"The CADx system described in this study was developed using a DenseNet-121 architecture trained on 85,000 annotated white-light and NBI colonoscopy images (see Methods: Model Development). Separately, during manuscript preparation, the authors used ChatGPT (GPT-4, OpenAI) to improve the language of the Discussion section. The research AI system and the writing assistance tool are entirely separate. All text edits were reviewed by the senior author (R.S.). The authors take full responsibility for the published content."

Microbiome research

Gut microbiome studies represent a major content category for Gastroenterology. 16S rRNA gene sequencing, shotgun metagenomics, metabolomics, and microbiome-host interaction analyses all involve substantial bioinformatics. If AI coding assistants helped you write analysis scripts, that needs disclosure.

The analytical tools, QIIME2, MetaPhlAn, HUMAnN, PICRUSt2, are standard research software. They don't require AI disclosure. But if ChatGPT or GitHub Copilot helped you write the Python or R scripts that call these tools, that's AI-assisted code generation.

Example for a microbiome paper:

"Microbiome analysis was performed using QIIME2 (v2024.2) and MetaPhlAn 4 as described in Methods. GitHub Copilot (Microsoft) assisted with writing custom R scripts for differential abundance analysis and visualization. All scripts were validated against published benchmark datasets. Claude (Claude 3.5, Anthropic) was used to improve the readability of the Introduction and Results sections. The authors take full responsibility for the published content."

IBD clinical trials

Gastroenterology is a primary destination for inflammatory bowel disease clinical trials, new biologics, small molecules, combination therapies. These papers influence treatment algorithms that gastroenterologists worldwide follow. AI involvement in interpreting treatment response rates, endoscopic endpoints, or histologic scores would undermine the clinical credibility of the paper.

The same principle applies to hepatology endpoints that sometimes appear in Gastroenterology: fibrosis staging, cirrhosis assessments, liver function outcomes. AI can help you say things more clearly. It shouldn't help you decide what to say about clinical outcomes.

AGA clinical guidelines

Gastroenterology publishes AGA Clinical Practice Guidelines and Technical Reviews. These documents directly guide gastroenterologist decision-making for conditions like GERD, IBD, Barrett's esophagus, and colorectal cancer screening. The AI sensitivity here is maximal. While the formal policy doesn't create separate rules for guidelines, the editorial expectation is that AI tools play zero role in guideline content generation.

For an original research article (e.g., IBD cohort study):

"During the preparation of this manuscript, the authors used ChatGPT (GPT-4o, OpenAI) to improve the clarity and readability of the Discussion section. All AI-generated suggestions were reviewed and edited by the corresponding author (T.U.) and the senior gastroenterologist (V.W.). No AI tools were used in study design, data collection, statistical analysis, or interpretation of clinical outcomes. The authors take full responsibility for the content of this article."

For a randomized controlled trial:

"The authors used Claude (Claude 3.5, Anthropic) to improve the English language of the Introduction and Methods sections. No AI tools were used for trial design, data management, statistical analysis, endpoint assessment, or clinical interpretation. All efficacy and safety conclusions were drawn by the principal investigators based on the pre-specified analysis plan. The authors take full responsibility for the published content."

For a technical review or meta-analysis:

"During the preparation of this technical review, ChatGPT (GPT-4, OpenAI) was used to improve sentence-level readability of the evidence synthesis sections. The systematic literature search, study screening, data extraction, risk of bias assessment, and evidence grading were performed entirely by the author team without AI assistance. The authors take full responsibility for the published content."

These disclosures all share the same DNA: specific tool identification, clear scope description, explicit statement of what wasn't AI-assisted, and a responsibility declaration. You'll notice that the clinical papers go out of their way to clarify that AI didn't touch the clinical interpretations. That's intentional and it's what Gastroenterology's editors expect.

Consequences follow the standard COPE-based escalation that most major journals use, but with the added dimension of Elsevier's institutional infrastructure:

During review. If a reviewer or editor suspects undisclosed AI use, you'll receive a query. This doesn't guarantee rejection, but it does change the editorial relationship. Your paper moves from "under normal review" to "under scrutiny," and that's not a helpful shift.

After acceptance. Discovery during the production phase halts publication. You'll need to add a proper disclosure and explain the omission. Elsevier's editorial system tracks these incidents, and the record follows the manuscript through any future correspondence.

After publication, the escalation ladder:

1. Published correction. For cases where AI use was limited to language editing and didn't affect scientific content, a corrigendum adding the disclosure may be sufficient.
2. Expression of concern. If the AI use raises questions about the reliability of the scientific findings, for instance, if AI was used to draft clinical interpretations or synthesize literature, the editor may issue an expression of concern.
3. Retraction. Reserved for cases where AI involvement was extensive enough to undermine confidence in the paper's conclusions. Rare, but possible.
4. Elsevier Research Integrity involvement. Serious cases may be escalated to Elsevier's central Research Integrity team, which can investigate across the publisher's entire journal portfolio. If they find a pattern of undisclosed AI use by the same author across multiple Elsevier journals, that's a bigger problem than one missed disclosure.
5. Institutional notification. The journal may contact the corresponding author's institution, triggering a formal integrity investigation.

The combination of AGA's clinical reputation and Elsevier's publishing infrastructure means that integrity issues at Gastroenterology get attention from both the clinical community and the publishing industry. That's more exposure than most authors want for an ethics violation.

Key observations from this comparison:

Gastroenterology and Gut have similar rules but different organizational structures. Gastroenterology's AGA/Elsevier model and Gut's BMJ Publishing Group model produce essentially equivalent author requirements. The practical differences are in submission systems and editorial workflows, not in what you're allowed or required to do.

Lancet Gastroenterology & Hepatology stands out for two reasons: it places the AI disclosure in the acknowledgments (not Methods), and it applies the Lancet's stricter interpretation that limits AI to "readability and language" improvements only. If you're considering Gastroenterology versus Lancet Gastro for a GI paper, note that the Lancet title is more restrictive on what AI can do.

Hepatology and the American Journal of Gastroenterology both use Wolters Kluwer as their publishing partner but take their policy direction from their respective societies (AASLD and ACG). The result is similar to Gastroenterology's model: society sets the clinical standards, publisher provides the infrastructure.

All five journals agree on the core rules. No AI authorship, mandatory disclosure, no AI-generated images, full author responsibility. The GI field has reached consensus on fundamentals even as the organizational structures differ.

The Elsevier submission system

Gastroenterology uses Elsevier's Editorial Manager for manuscript submission. During the upload process, you may encounter specific questions about AI use. Don't skip these or click through them reflexively. Your answers become part of the formal submission record, and inconsistencies between what you declare in the system and what's in your manuscript will be noticed.

Multi-center GI studies

Large multicenter GI studies, IBD registries, screening colonoscopy databases, NAFLD cohorts, can have dozens of co-authors across institutions in different countries. Each co-author may have different habits with AI tools. The corresponding author needs to actively survey all co-authors about AI use before finalizing the disclosure statement.

This isn't just good practice; it's a requirement. If a co-author used AI tools without telling you, and this comes to light after publication, the corresponding author bears primary responsibility for the incomplete disclosure.

Supplementary materials

Many Gastroenterology papers include extensive supplementary materials: additional tables, supplementary figures, extended methods. The AI disclosure requirement covers supplementary content too. If AI helped edit your supplementary methods section, that's covered. Don't assume the disclosure only applies to the main manuscript text.

Don't over-think the exemptions

Standard spell checkers don't need disclosure. DeepL for basic translation of a sentence doesn't need disclosure. But DeepL Write, Grammarly's AI rewriting features, and any tool that generates or substantially rewrites text does need disclosure. When in doubt, disclose. There's no penalty for disclosing something that didn't strictly require it. There is a penalty for not disclosing something that did.