AI Slop in Science Papers: Practical Rules for Authors

Updated 5 November 2026 · Standard formula, shown below

Fighting AI slop in science papers requires authors to verify every claim and data point, editors to set transparent policies and proportionate checks, and reviewers to focus on evidence rather than trying to guess who used a tool. Generative AI can assist with limited tasks, but it cannot take responsibility for research; disclosure, human verification and ordinary research-integrity safeguards are essential.

What “AI slop” means in science

“AI slop” is an informal label, not a technical or regulatory category. In scholarly publishing it usually describes low-quality or misleading material produced or polished with generative tools: fabricated references, unsupported claims, generic prose, incorrect summaries, invented data, or images that misrepresent evidence. The same problems can also arise without AI, so a suspicious style alone is not proof of misconduct.

Generative systems produce plausible sequences of words or images; they do not reliably distinguish true statements from false ones. A fluent paragraph can contain a nonexistent citation, reverse the findings of a paper, or turn a tentative association into a causal claim. Errors are particularly hard to spot when the subject is outside an author’s expertise or the prose sounds confident.

Not every use is harmful. A researcher might use a tool to improve grammar, translate a draft, or help organize code. The important distinction is whether the use preserves accuracy, confidentiality, authorship accountability and the integrity of the research record.

Practical rules for authors

Authors remain responsible for the complete submission, including content generated or edited with software. Before submission, use this checklist:

Verify each factual claim. Trace important statements to primary sources, and check that the cited work actually supports the claim. Open every reference and confirm its title, authors, journal or venue, year, and relevant page or identifier.
Check data and analysis at the source. Compare reported numbers with original records, code, tables and figures. Re-run analyses where possible. Never treat generated examples, synthetic data or suggested results as observations from an experiment.
Inspect every figure. Confirm that images have not been altered in a way that changes meaning. Disclose permitted processing, retain original files, and follow the journal’s figure-integrity rules. A generated illustration should not be presented as an experimental image.
Disclose relevant assistance. Follow the target journal’s policy and state the tool’s name, what it was used for, and, where requested, the version or date. A concise statement might say that a named tool assisted with language editing and that the authors reviewed and revised the text. Do not imply that a tool is an author.
Protect confidential material. Do not paste unpublished manuscripts, identifiable participant information, confidential peer-review content or restricted data into an external service unless the relevant permissions and institutional rules allow it.
Keep an audit trail for consequential use. Preserve prompts, outputs, code changes or other records when they materially shaped methods, analysis or results, especially if a journal or funder requests them.

Disclosure is not a substitute for verification. Conversely, a journal may not require disclosure for routine spelling correction; authors should check its current instructions rather than assume every tool use is treated alike.

Rules for editors and publishers

Editorial policies work best when they distinguish assistance from responsibility. State what uses are permitted, what must be disclosed, how confidentiality is protected, and whether generated text, images or code require special documentation. Make the rules easy to find before submission, and apply them consistently across disciplines and languages.

Editors should not rely on automated “AI detectors” as proof that a manuscript was generated by a model. Detector scores can be wrong and may vary with writing style, language background, editing, or text length. A score can at most prompt careful human assessment under a transparent process; it cannot establish authorship or misconduct on its own.

Use checks that address the actual risk. Reference verification, image-integrity review, statistical scrutiny, plagiarism screening and requests for underlying data or code can reveal problems regardless of how a passage was written. If concerns arise, ask specific, neutral questions and give authors a chance to explain or correct them. Follow established publication-ethics procedures for serious concerns rather than making a public accusation based on stylistic impressions.

Editorial teams should also protect submitted work. Manuscripts and reviewer reports are confidential unless the journal’s process says otherwise. Staff should not upload them to external tools without authorization and appropriate safeguards. Journals need to review their own workflows, disclose relevant uses of automation, and retain a human decision-maker accountable for editorial judgments.

What reviewers should—and should not—do

Peer reviewers evaluate the work, not an imagined author’s writing habits. Check whether the methods support the conclusions, whether references are genuine and relevant, whether figures match their descriptions, and whether reported analyses can be understood or reproduced. Flag a concrete discrepancy, such as a citation that does not exist or a result that conflicts with a table, rather than saying only that prose “sounds like AI.”

Reviewers should follow the journal’s confidentiality and tool-use rules. If a journal has not permitted external processing, do not submit the manuscript or its contents to a public chatbot or other service. Reviewers should not use generated summaries as a substitute for reading the paper and checking its evidence. If a potential integrity issue remains, report it confidentially to the editor, who can decide what evidence or clarification is needed.

Common failures and useful safeguards

The most frequent mistake is treating polished language as evidence of reliable scholarship. Other warning signs include references that cannot be found, citations attached to claims they do not support, inconsistent numbers across text and tables, unexplained shifts in terminology, and images that lack a credible source. These are reasons to investigate, not automatic proof of AI use.

Another misconception is that disclosure alone makes any use acceptable. A disclosed tool can still introduce false claims, expose sensitive information or distort an image. The reverse misconception—that any AI assistance is misconduct—is also too broad. Policies differ, and routine language editing may be allowed when the authors retain control and responsibility.

Practical safeguards are often more dependable than trying to identify a tool from prose:

Use reference managers and verify citations against the source, not just a generated bibliography.
Compare manuscript tables with analysis outputs and retain scripts and source data as required.
Ask a qualified coauthor to check technical sections outside the drafter’s expertise.
Use established image and plagiarism checks as screening aids, followed by human review.
Correct errors promptly and transparently; do not quietly replace questionable content after publication.

These controls address ordinary human error as well as AI-assisted error. They also preserve a useful distinction: the relevant question is whether the paper is accurate, ethical and accountable—not whether a detector assigns it a particular score.

Frequently Asked Questions

Can AI be listed as an author of a scientific paper?
No. A tool cannot take responsibility for the work, approve the final manuscript or respond to integrity concerns. Human authors must meet the journal’s authorship requirements and remain accountable.

Do authors have to disclose AI-assisted editing?
It depends on the journal’s current policy and the kind of assistance. Check the instructions for authors and disclose uses that the policy requires, describing the task clearly rather than making a vague statement.

Can an AI detector prove that a paper was generated by AI?
No. Detector output is not reliable proof of authorship or misconduct. Any concern should be assessed using specific evidence and a fair editorial process.

Is using AI to find references safe?
It can help suggest search terms or possible sources, but suggestions must be checked in reliable bibliographic records and against the original papers. Do not cite a source that you have not verified.

Should reviewers paste a manuscript into a chatbot?
Not unless the journal explicitly permits that use and confidentiality requirements are met. Reviewers should follow the journal’s policy and assess the manuscript’s evidence themselves.

What should an author do after finding an AI-generated error?
Correct the submission before publication and tell the editor if the error affects claims, data, methods or figures. For a published paper, contact the journal promptly and follow its correction process.

Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.