Could authors slip hidden instructions for AI reviewers into a review copy, then scrub them out before publication?

This explores whether authors could slip hidden instructions aimed at AI reviewers into the review copy of a paper and then delete them before publication. The collection doesn't study that exact switch, but it shows every piece needed for it to work. Hidden prompts are already in real papers. A Nikkei Asia investigation found invisible text in papers from at least 14 institutions in eight countries. The text was white-on-white or tiny, so it didn't show up in a normal PDF reader, but an AI that read the full document would see it and be told to write a flattering summary Are researchers hiding prompt injections in academic papers?. A separate audit found 18 arXiv manuscripts with hidden instructions telling AI reviewers to give positive assessments Are hidden AI prompts in preprints a deceptive research practice?.

The detail that answers your question is how these were caught: they were still in the public versions. That suggests the authors who were caught didn't bother removing them. A more careful author who put the prompt only in the submitted copy would leave nothing for a later scan of public papers to find. The arXiv audit also makes an ethics point that applies directly. The authors call it a questionable research practice because the text is both concealed and self-serving, whatever the author claims they meant by it Are hidden AI prompts in preprints a deceptive research practice?. Removing the prompt before publication would be two layers of concealment, not a defense.

The bigger question is whether hidden prompts are even needed. One study found that AI reviewers fail the two basic conditions for automating peer review. They agree with each other more than human reviewers do, a 'hivemind' effect. And they are easy to game: simply rewriting a paper's text with an AI raised AI review scores by about 0.45 points without improving the science at all Can AI systems safely replace human peer reviewers?. Hidden prompts are just the crudest version of a broader weakness. If AI reviewers respond to how a paper is worded, someone can steer them through what's visible as well as what's hidden. And because the reviewers tend to agree, one trick that works on one may work on all of them.

From other parts of the collection, two kinds of defense appear, though none was built for this problem. The first is to make AI review more rigorous. An agentic reviewer that checks proofs and experiments line by line caught serious flaws that had passed human review at major conferences Can inference scaling help reviewers catch errors humans miss?. At ICLR 2025, AI was used to give feedback to human reviewers, and 27% of them revised their reviews; the human stayed the final judge Can LLM feedback help peer reviewers improve their own reviews?. The second is to make documents traceable. Work on AI 'skills' argues that anything shaping an AI's behavior should be kept as versioned files that can be inspected and rolled back, not as hidden prompt state Can person-grounded skills remain auditable without hidden prompt state?. Applied to peer review, the fix for insert-then-remove is to archive the exact file reviewers saw and compare it with the published version. Writers in shared editors also said they want to see when and how AI was used Do writers want to see each other's AI prompts in shared editors?, which points toward disclosure as a norm.

What you might not have expected: the risk isn't only that a trick slips through. Sakana AI's fully AI-generated paper cleared a workshop's double-blind review with a 6.33 score, and its authors later found a citation error in it Can AI-generated papers pass peer review undetected?. The bigger danger is a review process where AI writes the papers, AI reviews them, and the review copy differs from the published one. In that setup, nobody is checking the version that actually matters.