Hidden AI prompts in academic papers manipulate peer review
Researchers at major universities in South Korea, Japan, the US, and China embedded hidden text prompts in academic papers posted on arXiv to manipulate AI-powered peer reviewers into providing positive evaluations. At least 17 papers were discovered containing such commands, written in invisible fonts to evade detection. A KAIST researcher acknowledged the practice was inappropriate and committed to withdrawing the paper.
What happened
Researchers at universities across at least eight countries embedded hidden text prompts in academic papers posted on arXiv to manipulate AI-powered peer reviewers. The hidden commands instructed AI systems to provide positive evaluations of the papers while ignoring criticisms, written in extremely small fonts or white text on white backgrounds to evade human detection.
At least 17 papers were identified containing such commands, primarily in computer science fields. Institutions identified included KAIST and Waseda University, as well as the University of Washington, Columbia University, Peking University, and National University of Singapore.
Reported impact
A KAIST paper co-author acknowledged that attempting to prompt AI peer reviewers for positive evaluations was inappropriate and committed to withdrawing the published paper. A Waseda University professor claimed the hidden prompts were intended as a countermeasure against lazy reviewers who rely solely on AI for paper evaluation.
KAIST's public relations office stated the practice was content it had not been aware of and announced plans to establish guidelines for appropriate AI use in research.
Index analysis
The incident revealed a vulnerability in peer review systems where AI tools can be manipulated through embedded prompts in the very documents they are evaluating. The tactic exploits the capacity of language models to follow instructions embedded in text without human oversight during automated review processes.
Nikkei Asia reported that journals and academic conferences have not yet established unified rules governing AI use in peer review, creating ambiguity about acceptable practices.
Governance controls
KAIST announced it would establish appropriate guidelines for AI use in research following the discovery. No evidence of systematic oversight or pre-submission audits to detect such manipulation was reported.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.