Scientists Criticize NeurIPS for Using Hidden Prompts to Catch AI-Assisted Peer Reviews
Key Takeaways
- ▸NeurIPS and ICML embedded concealed prompt injections in papers to detect peer reviewers inappropriately using AI tools to complete reviews
- ▸The hidden prompt technique has sparked controversy, with critics arguing it presumes bad faith and damages the trust between conferences and reviewers
- ▸ICML successfully identified hundreds of inappropriate AI uses in peer review through this method, resulting in rejection of ~2% of submissions
Summary
NeurIPS, the prominent annual machine learning conference, has embedded concealed prompt injections into submitted papers to detect when peer reviewers use generative AI in violation of the conference's policies. The hidden instructions are designed to produce telltale phrases in peer reviews that would indicate large language model (LLM) use, such as 'This work addresses the central challenge.' However, the technique has sparked significant backlash from the research community, with multiple scientists arguing it presumes bad faith and undermines trust in the peer review process.
Similar detection efforts at ICML 2026 have successfully identified hundreds of reviewers misusing AI tools, according to Carnegie Mellon computer scientist Nihar Shah, who chairs ICML's scientific integrity committee. The conference desk-rejected approximately 2% of total submissions—nearly 500 papers—over LLM review policy violations. While Shah and others defend hidden prompts as viable enforcement tools, critics like Sören Auer from Leibniz University Hannover argue the approach treats reviewers as suspects and corrodes the collaborative relationship essential to peer review. Researchers have also flagged concerns that reviewers may incorrectly penalize papers containing the hidden prompts, unaware they were inserted by conference organizers rather than paper authors.
- The debate reveals deeper tensions in the scientific community about how to regulate AI use—enforcement tactics versus transparent policies and cultural change
Editorial Opinion
While conference organizers face legitimate pressure to prevent AI misuse in peer review, the hidden prompt approach represents a troubling shift from building trust to setting traps. Treating reviewers as suspects rather than partners in quality assurance risks damaging the peer review system the conferences are trying to protect. The research community would benefit more from transparent policies on AI's appropriate role in peer review, updated workflows that enable beneficial tool use, and a culture that supports researchers in engaging with AI openly rather than in shadows.
