Science (AAAS) AI Policy: From Total Ban to Mandatory Disclosure

Science's current AI policy has four components:

1. AI can't be an author. Like every major journal, Science doesn't allow AI tools to be listed as authors. The reasoning: authors must take accountability for published work, and AI can't do that.

2. All AI use must be disclosed in three places. This is where Science differs from most journals. You must describe AI use in:

• The cover letter (so editors know before review begins)
• The acknowledgments section (so it's visible in the published paper)
• The methods or supplementary materials (so the technical details are documented)

Three disclosure points is more than Nature requires (Methods only) or Cell Press requires (dedicated section before References). Science wants AI use flagged at every stage of the editorial process.

3. AI-generated images require editor permission. You can't include images created by generative AI tools without explicit advance approval from your handling editor. This is slightly less absolute than Nature's outright ban, but in practice, getting that approval for a standard research paper would be unusual.

4. Violations constitute scientific misconduct. This is the sharpest teeth in Science's policy. Undisclosed AI use isn't treated as a paperwork error or an oversight. It's classified as misconduct, the same category as data fabrication and plagiarism. Misconduct findings can trigger retraction, institutional notification, and publishing bans.

Most journals treat AI policy violations as ethical lapses to be handled on a case-by-case basis. Science is more categorical. When Science says "scientific misconduct," it invokes a specific framework:

• The journal can notify your institution's research integrity office
• Your institution may be obligated to investigate under federal research misconduct policies (especially if you receive NIH, NSF, or other federal funding)
• A misconduct finding goes on your record and can affect future funding and publication

This doesn't mean Science will retract your paper if you forgot to mention you used Grammarly. The policy targets substantive undisclosed use, like generating entire sections of text without disclosure. But the classification signals that Science takes this more seriously than a typical policy violation.

Science's three-point disclosure requirement means you need to think about AI use before you even begin the submission process.

In the cover letter:

"During preparation of this manuscript, we used Claude (Anthropic) to assist with language editing of the Results and Discussion sections. All AI-assisted text was reviewed, verified, and edited by the authors, who take full responsibility for the content."

In the acknowledgments:

"The authors acknowledge the use of Claude (Anthropic) for language editing assistance during manuscript preparation."

In the methods/supplementary materials:

"Language editing: Sections of the manuscript were refined using Claude (Anthropic, Claude 3.5 Sonnet) for clarity and grammar. The authors reviewed all AI-assisted text and verified its accuracy against the original data and analysis."

The key principle: be specific about which tool, which version if relevant, and which parts of the manuscript were affected.

Science's image policy is worth understanding because it's slightly more nuanced than a blanket ban. The journal prohibits AI-generated images by default but allows exceptions with explicit editor permission. In practice, this means:

• Banned without permission: DALL-E, Midjourney, or Stable Diffusion graphics used as figures, schematics, or graphical abstracts
• Requires disclosure: AI-enhanced processing of real images (denoising, super-resolution, etc.)
• Potentially allowed with permission: AI-generated visualizations that serve a scientific purpose and are clearly labeled as AI-generated

The "with permission" pathway exists primarily for papers about AI image generation itself, or for computational studies where AI-generated visualizations are part of the research methodology. For a standard biology, chemistry, or physics paper, don't expect the editor to approve AI-generated figures.

Like most journals, Science relies primarily on author self-declaration rather than automated detection. The submission system includes checkboxes and attestations related to AI use. Editors and reviewers may flag suspicious text, but there's no systematic AI-detection screening.

Post-publication enforcement is where the misconduct classification becomes relevant. If another researcher, a reviewer, or a reader identifies undisclosed AI use after publication, Science has the policy infrastructure to treat it as a formal misconduct case rather than just issuing a correction.

Science's three-point disclosure requirement and misconduct classification make it the most demanding of the top-tier journals on AI transparency. Whether this is appropriate vigilance or excessive bureaucracy depends on your perspective, but either way, you need to comply.

Before writing:

• Decide upfront whether you'll use AI tools. If yes, plan your disclosure language early.
• Keep records of which tools you used and which manuscript sections they touched. You'll need this specificity for the three disclosure points.

During writing:

• Use AI tools if they genuinely help, but don't use them just because they're available. Every AI use creates a disclosure obligation.
• Don't have AI generate your cover letter, the document where you're supposed to disclose AI use. The irony won't be lost on editors.

During submission:

• Include AI disclosure in all three required locations. Missing even one creates a compliance gap.
• Be honest about the extent of AI involvement. "Used for minor language editing" when AI generated entire paragraphs is the kind of discrepancy that becomes a misconduct case.

Common pitfalls:

• Forgetting the cover letter disclosure. Most authors remember the manuscript sections but overlook that Science also requires it in the cover letter.
• Using generic disclosure language. "AI tools were used during preparation" isn't specific enough. Name the tool, describe the use.
• Assuming code generation doesn't count. If you used GitHub Copilot or similar tools for analysis code, that falls under the policy too.

Not sure if your manuscript meets Science's disclosure requirements? A free manuscript review can help you identify gaps before submission.

Science's AI policy applies to all journals in the Science family:

• Science
• Science Advances
• Science Translational Medicine
• Science Signaling
• Science Immunology
• Science Robotics

The same three-point disclosure requirement and misconduct classification apply across all six journals. If you're submitting to Science Advances thinking the rules might be softer, they aren't.