Journal AI Policies in 2026: What Authors Need to Know Before Submission

The high-level rule sounds similar everywhere: disclose AI use and keep humans accountable. The practical differences are more annoying than that:

• Nature Portfolio sorts AI use into green, amber and red tiers. Green-tier uses are permitted, with disclosure encouraged; assigning authorship to AI and creating photorealistic AI images are not permitted. The page asks for disclosure without naming a manuscript section.

• Elsevier asks for a separate AI declaration that appears in the published article, a description in Methods when AI was part of the research, and a caption note on AI-assisted figures. Spelling and grammar checkers are exempt, and AI may not create or alter primary research images such as blots or micrographs.

• JAMA Network puts AI help with writing in the Acknowledgment section and AI used in the research in Methods, and asks for the tool name, version, manufacturer, dates of use and what it did.

• ACS wants all AI use disclosed in the submission: a note in the Acknowledgments, fuller detail in Methods for substantial use, and a caption note on AI graphics. AI images are not allowed in Table of Contents graphics, which have no caption.

• IEEE asks authors to name the AI system, and the sections it produced, in the Acknowledgments. The rule covers text, figures, images and code; editing and grammar help falls outside it.

• PLOS wants the tools, what they were used for and how you checked the output, in Methods, Acknowledgments or the figure legend. Language polishing needs no disclosure, but generated code does.

That is why generic "we used AI for editing" language is often too weak. A compliant statement needs the tool, the task, and the human review step.

In our pre-submission review work, the authors who get into trouble are usually not the ones doing wild things with AI. They are the ones assuming all publisher policies are basically the same and dropping in vague disclosure language at the last minute.

The practical mistake is simple: a paper gets retargeted across publishers, but the disclosure does not get retargeted with it. That is how a statement that looked acceptable at one journal becomes incomplete or misplaced at the next one.

Every journal's policy is different in the details, but these five steps cover the compliance requirements across all major publishers:

• Check your target journal's specific AI policy. Policies are not standardized. Some journals require disclosure in methods, others in acknowledgments, others in a separate submission form field. Check the author guidelines for your specific target journal, not just the publisher umbrella page.

• Document which AI tools you used and how. Be specific: "We used ChatGPT (OpenAI, GPT-4) to assist with language editing of the discussion section. All output was reviewed and revised by the authors." That level of detail (the tool, the task, the human check) matches what most of the policies in the tracker ask for. "We used AI to help write the paper" is too vague.

• Verify every citation and factual claim AI touched. AI tools hallucinate citations, fabricate statistical claims, and generate confident-sounding statements that are factually wrong. A 2025 analysis found over 100 hallucinated citations in papers accepted at a top machine learning conference. Every reference must be verified against the actual source. A citation verification scan catches fabricated references before reviewers do.

• Don't use AI-generated images without disclosure. Some journals explicitly prohibit AI-generated figures unless disclosed and justified. If you used DALL-E, Midjourney, or similar tools to create any visual content, check whether your target journal permits it.

• Don't claim AI-generated text as entirely your own. If an AI tool wrote substantial portions of the manuscript, this must be disclosed. Not disclosing is a form of plagiarism under most journal policies. The risk is not just rejection but retraction and reputational damage if discovered post-publication.

Abstract policy language doesn't convey the real consequences. These are composites of cases that integrity committees have dealt with.

The hallucinated reference list. A biomedical research team used ChatGPT to draft their introduction and let it suggest supporting citations. Twelve of forty-three references didn't exist, plausible-looking combinations of real author names, real journals, and fabricated titles. Two reviewers independently flagged the issue. The manuscript was rejected, the corresponding author received a formal warning, and the flag went on the author's submission record.

The AI-generated figure with visual artifacts. A group submitted a review article with AI-generated schematic diagrams containing nonsensical text labels, impossible anatomical geometry, and repeating noise patterns. The production team caught it during typesetting. Complete figure replacement plus a six-week publication delay.

The discussion section that said nothing specific. A reviewer returned a manuscript with a single comment: "This reads like it was generated by a language model. Every paragraph makes broad claims without connecting to the specific results of this study." The paper eventually published, but only after two additional revision rounds.

The principle: AI accelerates each step, but a human must own the output at every stage.

The safe rule for drafting: write methods and results yourself (these require domain precision AI can't match), then use AI for language polishing on the introduction and discussion. Rewrite anything that sounds generic.

Always disclose AI use if:

• The tool drafted, rewrote, or substantially edited any manuscript text (Elsevier, Cell Press and AAAS all ask for this, and Nature Portfolio encourages disclosure even for its lowest-risk uses)

• AI generated or suggested any references, and you've verified every one actually exists

• AI wrote or changed code used in the analysis (PLOS and IEEE both name code in their rules)

• You used AI for figure creation, data visualization, or image generation (some journals prohibit this entirely)

• Your target journal has an AI checkbox or text field in the submission portal (skipping it is a procedural flag)

• You're unsure whether your use counts, the cost of disclosing is zero, the cost of not disclosing is retraction

You don't need to disclose if:

• You used standard spell-check or grammar tools that aren't AI-powered (e.g., built-in Word spell-check)

• Your only AI use was spelling or grammar help at a publisher that exempts it: Elsevier, Wiley, Sage, PLOS and MDPI exempt it, and IEEE recommends but does not require disclosure

• The journal's policy explicitly excludes the type of use you made (read the actual policy, not the publisher umbrella page)

When policies conflict across journals: If you're resubmitting a rejected paper to a different publisher, rewrite the disclosure to match the new journal's format. Elsevier wants a separate AI declaration printed with the article. JAMA Network wants writing help in the Acknowledgment section. Nature Portfolio asks for disclosure without naming a section, so follow the journal's own instructions. Do not reuse the old disclosure unchanged. Adapt it.