International computer science researchers are embedding hidden text in academic papers to influence AI review outputs. These prompt injection attacks have been observed across 14 academic institutions in eight countries. The text, designed to be invisible, instructs AI models to focus on positive comments. An inquiry into English language preprints on ArXiv led to the identification of 17 such papers. Some contained explicit instructions to reviewers to ignore negative feedback and provide positive evaluations. One paper intended for an upcoming conference is being withdrawn amid these revelations.
A handful of international computer science researchers are attempting to influence AI reviews through prompt injection attacks, embedding secret instructions in academic papers.
Hidden text styled to be invisible has been found in research papers that instruct AI models to provide flattering summaries, affecting reviews in the AI landscape.
Collection
[
|
...
]