AI Detection in Academic Publishing and Peer Review

Academic publishing runs on trust: that the authors did the work, that the references exist, and that reviewers read the paper themselves. Generative AI tests all three.

Much AI use in research is legitimate, especially language help for researchers writing in their second language. The goal for editors and research offices is not to ban AI. It is to make sure its use is disclosed, that humans stay accountable, and that invented content does not slip through.

Try it: use ProofAI's PDF AI checker, free for 10 checks a day with no login.

How common is AI-assisted writing in research?

A study in Nature Human Behaviour by Weixin Liang and colleagues analysed 1,121,912 papers and preprints from arXiv, bioRxiv and 15 Nature portfolio journals, published between January 2020 and September 2024. By September 2024, they estimated that about 22.5% of sentences in computer science abstracts showed signs of substantial LLM modification, compared with 7.7% in mathematics abstracts and 8.9% in Nature portfolio abstracts.

Two caveats from the authors matter. The method estimates a share across a whole body of papers; it does not classify individual papers. And it measures substantial modification, not proof of misconduct.

Peer review is affected too. An ICML 2024 paper by the same research group estimated that 6.5% to 16.9% of review text at four AI conferences in 2023 and 2024 had been substantially modified by LLMs. Estimates were higher for reviews submitted within three days of the deadline.

The risks editors actually face

Invented references. In a 2023 Scientific Reports study, Walters and Wilder checked 636 citations in 84 short papers generated by ChatGPT. They found 55% were fabricated with GPT-3.5 and 18% with GPT-4. Newer models make fewer errors, but every reference still needs checking.

Want to check your own content? Try ProofAI free 10 checks a day, no login needed.

Undisclosed AI text. Most journal policies allow some AI use if it is declared. The problem is AI-written content presented as the authors' own.

Confidentiality breaches. A reviewer who pastes an unpublished manuscript into a public AI tool may expose it.

Also read: India's AI Labelling Rules 2026: What the IT Rules Require

Unfair accusations. Detectors can be wrong. A Stanford study by Liang and colleagues found that seven widely used detectors flagged human-written TOEFL essays by non-native English writers as AI-generated at an average rate of 61.22%, while classifying essays by US eighth-graders almost perfectly. For journals with many authors writing in a second language, a detector score alone must never be grounds for rejection.

What the main guidelines say

COPE. The Committee on Publication Ethics says AI tools cannot be listed as authors, because they cannot take responsibility for the work, declare conflicts of interest or manage copyright. Authors who use AI for writing, images or data work should disclose how and which tool in the methods section or similar, and remain fully accountable for the content.

ICMJE. The ICMJE Recommendations (updated January 2026) ask journals to require AI disclosure at submission. Writing help goes in the acknowledgments; AI use in data collection, analysis or figures goes in the methods. Reviewers must ask the journal's permission before using AI on a review, and uploading a manuscript to an AI tool may breach confidentiality.

Individual publishers add their own rules, so always check the journal's policy.

Signs a manuscript needs a closer look

None of these prove AI use. They are reasons to look harder.

References that cannot be found, or whose details do not match the real paper
Leftover chatbot phrases, such as "As an AI language model" or "Certainly, here is"
A generic abstract that could describe many papers
Methods too thin to reproduce
Sudden shifts in style or terminology between sections
Figures with odd details, garbled labels or duplicated regions

Where AI checks fit in the workflow

At submission. Ask authors to declare any AI use in a standard form, following COPE and ICMJE.

At editorial screening. Spot-check references, starting with any you do not recognise. If you use an AI detector, use it to decide where to look, not to make a decision.

During peer review. Remind reviewers of confidentiality and the journal's AI policy. Ask them to focus on methods, data and conclusions.

Before a decision. If concerns remain, ask the authors specific questions: which tool was used, for which sections, and why it was not declared. Give them a chance to respond.

Advice for universities and research offices

Publish a clear AI-use policy that matches the main journal guidelines.
Train students and early-career researchers on disclosure, not just on what is banned.
Offer optional pre-submission checks so authors can review their own drafts.
Keep a documented process for integrity concerns, with a right to reply.

Using a detector as a screening aid

ProofAI's document detector accepts PDF, DOCX, DOC and TXT files up to 25 MB and highlights the passages that most influenced the score, so an editor can see which sections to read closely. The first 10 checks a day are free with no account, or 25 with a free account, and a REST API is available for submission systems. ProofAI's results are probabilities, not proof, and given the bias evidence above, they should always be weighed with human judgement and the author's explanation.