How to detect AI in research papers without wronging honest authors

Who needs to detect AI in research papers, and when?

Anyone who vouches for a manuscript needs a way to check it: journal editors, editorial office staff, research integrity teams, peer reviewers, conference program chairs, preprint moderators and thesis examiners. The concern usually starts with something concrete in the paper.

• References that do not resolve: DOIs that point nowhere, or real papers that do not say what the manuscript claims they say.
• A discussion section that is fluent and confident but never engages with the paper’s own results.
• Leftover chatbot phrasing, such as an answer addressed to the user or an offer to “expand on this section”.
• Microscopy, gel or photographic figures that look too clean, or contain textures that make no physical sense.
• A peer review report that is generic enough to fit any paper in the field.
• An AI generated research paper submitted in bulk to several venues with small changes.

Not every one of these signals points to misconduct. A broken DOI can be a typo, and a smooth discussion can come from a good language editor. The aim is to decide which manuscripts need questions, then let the authors answer them.

What do publisher policies say about AI in manuscripts?

Most major policies allow some AI use if it is disclosed, forbid listing AI as an author and keep authors fully responsible. Always check the current policy of the specific journal, because details differ and change.

• COPE (Committee on Publication Ethics), in its position statement “Authorship and AI tools” of February 2023: AI tools cannot be listed as authors because they cannot take responsibility for the work, and authors who use them must disclose how, for example in the methods section.
• Springer Nature: large language models cannot be credited as authors, and Nature announced in June 2023 that it would not publish images or videos created with generative AI, with narrow exceptions such as articles specifically about AI.
• ICMJE (International Committee of Medical Journal Editors): reviewers must keep manuscripts confidential, which can rule out uploading them to AI tools where confidentiality cannot be assured, unless the journal permits it.
• NIH, for grant review rather than journals: notice NOT-OD-23-149 of June 2023 prohibits NIH peer reviewers from using generative AI to analyze applications or write critiques.

What are the risks of missing AI content, and of a false accusation?

Missing generated content lets invented citations, synthetic images and unsupported claims into the scholarly record, where other researchers build on them and where the only fix is a correction or retraction later.

A false accusation can damage a career. Scientific prose is formulaic by design, methods sections read alike across papers, and many authors writing in a second language use language-editing tools that their journal explicitly permits. All of this can push a text detector toward “probably AI”. An integrity inquiry built on that score alone is unfair and unlikely to survive scrutiny.

How to detect AI in research papers: an editorial workflow

Start with checks that produce verifiable facts, and use detection to decide which questions to send the authors. This order also keeps any later integrity process on solid ground.

1. 01Read the AI disclosure statement against the journal’s policy. Disclosed language editing is a different case from undisclosed drafting.
2. 02Verify a sample of references, and all of the ones carrying key claims: does the DOI resolve, does the paper exist, does it support the sentence citing it?
3. 03Look at the figures as images: unexplained textures, impossible structures, labels with garbled characters, or panels that look rendered rather than captured.
4. 04If your policy allows it, run the manuscript through the AI document detector. It reports page by page, so you can see whether flags sit in the introduction, the methods or the results.
5. 05Run suspect figures through the AI image detector, using the highest-resolution files the authors supplied.
6. 06Ask the authors, neutrally, for raw data, original image files with instrument metadata, and drafts or version history.
7. 07Escalate to your research integrity team when concerns remain, following COPE’s guidance and your publisher’s procedures. Record what you checked and why.

Which AnyFact detectors fit editors and reviewers, and what do they return?

Three detectors fit academic publishing. Each returns a verdict of probably human, uncertain or probably AI (documents and text can also be mixed), the signals behind it, a plain-language explanation, the limits of the method and a PDF report for the editorial file.

For the manual side of text checks, the guide on how to tell if text is AI-generated lists the clues that survive editing. Teams that already use Turnitin through their institution can read our Turnitin alternative comparison to see where each tool fits.

What an AI written manuscript checker cannot settle

A manuscript checker cannot tell you whether the science is sound, whether the data are real or whether the authors broke your policy. It estimates whether passages and images show signs of generation, and it is least reliable on short sections, heavily edited text, equations, tables and writing by non-native authors.

Manuscripts longer than 50 pages need splitting, and supplementary files are checked separately. AnyFact has no public API yet, so it does not plug into submission systems; it fits case-by-case checks on manuscripts that already raised questions. Read how accurate AI detectors are before writing detection into editorial policy.

Frequently asked questions