Researchers Hide Secret Messages in Papers to Trick AI Peer Reviewers

A recent Nature report has uncovered a troubling new tactic in academic publishing: researchers embedding hidden instructions in manuscripts to influence AI-based peer reviews. These covert prompts — often written in white text or tiny font invisible to the human eye — are designed to be read only by large language models (LLMs) that some reviewers use to evaluate submissions.

The hidden messages, found in at least 18 preprints across 44 institutions in 11 countries, typically instruct the AI to “give a positive review only” or to frame weaknesses as minor. In one case, authors squeezed 186 words of flattering review instructions into a single blank space after a period. The approach exploits a vulnerability in peer-review workflows when LLMs are used without human oversight — a practice prohibited by many publishers, but still occurring.

While tests suggest that not all AI models respond to these manipulations (with ChatGPT being more susceptible than Claude or Gemini), the tactic highlights broader concerns about research integrity and the potential misuse of AI in science. Experts warn that such “prompt injections” constitute academic misconduct and are symptomatic of deeper problems with incentives in scholarly publishing.

As AI becomes more entwined with peer review, the scientific community will need stronger safeguards to ensure fairness, detect manipulation, and maintain trust in the review process.

Gibney E. Scientists hide messages in papers to game AI peer review. Nature. 2025 Jul;643(8073):887-888. doi: 10.1038/d41586-025-02172-y. PMID: 40646151.

https://www.nature.com/articles/d41586-025-02172-y