Scientists decry conference's use of hidden prompts to snare AI peer reviews

Summary

OLDER ITEM (1 Jul 2026) included because it records a concrete enforcement event the knowledge base is missing. NeurIPS 2026 embedded concealed canary instructions in submitted papers - directing any language model to insert set phrases such as "This work addresses the central challenge" - to catch reviewers outsourcing reports to LLMs; ICML 2026 ran the same trap and caught hundreds of violators, producing roughly 500 desk rejections. The reported dispute is the substance: Soren Auer's objection that "designing a trap that presumes bad faith corrodes the relationship the whole system depends on" against ICML's Nihar Shah defending it as necessary given the scale of misuse. This inverts the usual prompt-injection framing - here the venue, not the author, is the injector.

Contribution

Queued — not yet read.

Method

Queued — not yet read.

Relevance to RISE

Surfaced by the 2026-09-07 literature sweep as a medium-priority candidate for the RISE knowledge base (angle: AI in peer review). Verified against: Opened https://www.thetransmitter.org/publishing/scientists-decry-conferences-use-of-hidden-prompts-to-snare-ai-peer-reviews/ on 2026-09-07; confirmed headline, byline Dalmeet Singh Chawla, date 1 Jul 2026, and the reported facts (NeurIPS canary prompts, ICML 2026 ~500 desk rejections, Auer and Shah quotes). Not in dedupe_titles.txt (checked 'hidden prompts', 'transmitter', 'icml 2026').

Critique / open questions

Queued — not yet assessed.

Key quotes

—