Researchers Are Hiding AI Prompts in Their Papers (cacm.acm.org)

Researchers are covertly embedding “prompt injection” instructions in preprints to try to influence AI-assisted peer review. Nikkei Asia found 17 arXiv papers containing hidden prompts that tell models (e.g., ChatGPT, Claude) to produce favorable reviews; independent researchers report similar isolated cases, largely concentrated in computer-science communities. The technique works because AI tools used by busy reviewers will read any text in a submission, so short, obscured directives can bias automated summaries and critiques without alerting human reviewers. That threatens the integrity of peer review by exploiting opaque AI workflows and the rising reliance on generative tools. The episode highlights both a security and workflow problem: authors are gaming reviewers, reviewers are outsourcing work to imperfect AI, and journals lack standards to handle either. So far there’s little evidence this is widespread outside preprints, but detection is hard and the practice could grow as prompt-injection techniques spread. Proposed fixes include automated screening for malicious prompts, clear publisher policies on AI use, and — more fundamentally — retooling papers to be machine-readable: structured appendices with code, data and standardized metadata that let AI verify analyses (re-run code, check statistics) rather than be swayed by prose. Absent structural reforms to reduce volume and improve reproducibility, the field risks an arms race between manipulative authors and hardened review tools.