Your Next Reviewer Will Probably Use AI. Here's How to Submit a Manuscript That Survives Both.

A Nature survey of 1,600 researchers published late last year found that more than half of reviewers used AI tools while preparing a peer review, in many cases against their journal's stated guidance. A few months later, an analysis of ICLR 2026 submissions reported that 21% of the conference's peer reviews were AI-generated end-to-end, with more than half showing some AI use. If your manuscript is in review, there is a coin-flip chance an AI is reading it.

1 in 5 reviews AI-generated end-to-end. That is the ICLR 2026 rate, with more than half of all reviews showing some AI involvement.

Your manuscript now needs to hold up to two kinds of reader at once: a fast AI that scans for the obvious flags, and a human who handles the methodology and the novelty call. The two readers catch different things. Planning for one and ignoring the other costs you the submission.

#What AI Reviewers Actually Catch

The Nature reporting splits reviewer AI use into rough buckets. About 53% of reviewers have tried it for something. Around 19% use it to examine methodology, statistics, or the logic of the argument. The remaining use clusters on surface tasks, and surface tasks are where AI reviewers are strong.

Three patterns show up across editor reports and competing tool analyses.

Reporting checklists and formatting. AI is fast at confirming whether a manuscript hits the items on CONSORT, STROBE, PRISMA, or ARRIVE, whether ethics approval is stated, whether a data availability statement exists, and whether the structure follows the target journal's template. These are the most preventable desk rejections in academic publishing.

Language clarity and structure. Long sentences, unexplained acronyms, an abstract that buries the contribution, a discussion that restates the results. AI flags these at the paragraph level.

Citation integrity. Fabricated references now appear in 1 in 277 papers, and AI reviewers cross-check bibliographies against Crossref and Semantic Scholar faster than a human can. A hallucinated entry in your reference list is more likely to be caught by an AI reader than by a human who is skimming.

AI reviewers catch what is countable and what is lookup-able. That is most of what gets a paper desk-rejected, and almost none of what gets it rejected on the merits.

#Where AI Reviewers Fail

The ICLR 2026 controversy is useful because the failure modes played out in public. Editors and authors who flagged the AI-generated reviews pointed to the same red flags: hallucinated citations inside the review itself, verbose feedback that did not engage with the core contribution, and reviews that misread what the paper was claiming.

Three failure modes worth naming.

Novelty in cross-disciplinary work. AI reviewers anchor on adjacent literature in the obvious subfield. A manuscript that braids cardiology methods into a social science question, or applies an ML technique to a clinical dataset, is the kind of paper an AI reviewer reads as either derivative or out of scope. The novelty sits in the seam between fields, and the seam is what an AI reviewer compresses away.

Methodology depth. Power analysis, choice of estimator, handling of missing data, model assumptions, sensitivity checks. A human stats reviewer digs into these. An AI reviewer, per the PNAS analysis of how shallow most AI tool engagement runs, waves through methods that are reported even when they are wrong.

Confident wrongness. An AI review can read as detailed and authoritative while being off about what the paper says. That is the worst failure mode for an author, because the editor sees a long, organized review and is less likely to discount it.

Surviving an AI reviewer is a different exercise from surviving a human one. You need to plan for both.

#The Polish-Depth Paradox

A counterintuitive finding from the PNAS analysis of AI use in academic writing: submission writing quality has dropped about 1.28 standard deviations from January 2021 to January 2026, while AI tools have raised the median surface polish of every paragraph. The two facts are not in tension. AI lifts the floor on grammar, transitions, and sentence rhythm. It does not lift the floor on what a paragraph actually says.

For an author this changes what a clean-looking draft signals. A polished manuscript no longer telegraphs quality to a calibrated reviewer; it telegraphs wrote this with help. Reviewers, including AI reviewers, are reading polish as a baseline rather than a virtue. The remaining signals of quality are concrete claims, specific numbers, and an argument that holds together under a second pass. If your draft reads smoothly but you cannot summarize each section's contribution in one sentence, an AI reviewer will probably not catch the gap. A human reviewer will, and the editor will weight that catch heavily.

The practical move is to write to a stricter standard for substance than your surface polish alone would imply. The polish is no longer doing work for you.

#The Journal Policy Whiplash

Journal AI policies are moving faster than most authors realize. A cross-disciplinary analysis in Learned Publishing found that between March and August of 2025, 24.5% of high-impact-factor journals revised their AI peer review policies. The share of journals holding an explicit position rose from 77% to 83% in five months. The latest count for top medical journals puts explicit AI guidance at 78%, and the policies do not agree with each other.

A few journals require disclosure of any AI use in writing. Others prohibit AI use in peer review and ask authors to identify which sections used AI assistance and how. The policies are inconsistent enough that a one-size disclosure no longer travels.

Check your target journal's AI policy before you finalize the cover letter. Submitting a manuscript that violates a policy the journal posted six weeks ago is now a routine reason for rejection without review.

#The Asymmetric Disclosure Problem

Authors are expected to disclose AI assistance. Reviewers, in most cases, are not. The Nature reporting makes this asymmetry explicit: more than half of reviewers have used AI, most of that use happens without the editor or author knowing, and enforcement of the few reviewer-disclosure rules in place is rare.

This shows up in two places that matter for your submission.

The first is your AI disclosure statement in the cover letter. A specific disclosure ("we used X for copyediting in the introduction and discussion; no AI was used for analysis or interpretation") closes off a vector that editors are checking on rejection. A vague disclosure or no disclosure opens it.

The second is the rebuttal. If a review has AI hallmarks (verbose, off-target, hallucinated references), the editor has likely noticed too. Your rebuttal does not need to call this out. It needs to be precise where the review was vague. Quoting the manuscript line, citing the table or figure number, and pointing to the specific paragraph that already addressed the reviewer's concern shifts the editor's weight away from a sloppy review without forcing you to litigate its provenance.

#What to Change Before You Submit

Five concrete changes, ordered by how much friction they remove.

1. Run a self-review that mirrors what AI reviewers check first. Reporting checklist completion, citation integrity, language clarity, logical flow. These are the easy flags, and they cover most of what gets a paper desk-rejected. Fix them before any reviewer sees them.

2. Verify every citation. This is the new baseline. See our walkthrough of how hallucinated citations slip through and what verification catches.

3. Pre-empt the methodology critique a human reviewer will catch even when the AI does not. Most rejections on the merits trace back to a handful of recurring problems. We covered them in 5 Common Methodology Issues That Get Manuscripts Rejected.

4. Check the journal's AI policy and write your disclosure to match. A two-line statement that names the tools you used and the sections they touched is enough at most journals. A vague statement is worse than no statement.

5. Plan your response letter for an AI-flavored review. Long, verbose, sometimes confused reviews are now common. The point-by-point approach in our response letter guide still works, with one adjustment: when a reviewer comment is wrong about your paper, quote the manuscript line in your response instead of paraphrasing it.

If you want a second pass before submission, run your manuscript through ManuscriptMind and get the AI-reviewer flags in the same report as the human-style critique.

#How ManuscriptMind Helps

ManuscriptMind runs the surface checks an AI reviewer would run, the structural and methodology critique most AI tools skip, and a per-reference citation report, in a single pass. You see the issues an AI reviewer is likely to flag before submission, and the deeper issues a human reviewer is likely to flag after. The categorization (critical, major, minor) maps to how an editor will read the comments. Most manuscripts finish in under five minutes.

The intent is narrow. We do not want to write your paper. We want to give you the same view of the manuscript a careful reviewer will have, while you can still change it.

Run a pre-submission review →

---

Related: Hallucinated Citations Are a Sixfold Problem and How to Respond to Reviewer Comments Without Tanking Your Resubmission.