ICML 2026 Desk-Rejected 497 Papers After Reviewers Broke LLM Rules They Agreed To

ICML 2026 has taken decisive action against reviewers who violated the conference's LLM usage policies — desk-rejecting 497 papers (approximately 2% of all submissions) authored by reviewers who were caught using LLMs in their reviews despite explicitly agreeing not to.

The Two-Policy Framework

ICML 2026 offered reviewers a choice between two policies:

• Policy A (Conservative): No LLM use allowed in reviews at all
• Policy B (Permissive): LLMs allowed to help understand papers and polish reviews

Reviewers self-selected their policy. Critically, no reviewer who strongly preferred Policy B was assigned to Policy A — only those who explicitly chose "no LLMs" or "I'm okay with either" were placed in the restricted group.

The Violations

After assignment, 795 reviews (~1% of all reviews) from 506 unique Policy A reviewers were detected as LLM-generated. These were reviewers who had explicitly agreed to not use LLMs.

The detection method did not rely on generic AI-text detectors. Every flagged instance was manually verified by a human to avoid false positives.

The Consequences

The punishment was severe and targeted:

• 497 papers desk-rejected — the corresponding author's submission was rejected if their assigned reviewer produced LLM-generated content under Policy A
• All detected Policy A LLM reviews removed from the system
• 51 reviewers removed entirely from the reviewer pool — these were reviewers where more than half of their reviews were LLM-generated (10% of the 506 detected reviewers)

Why This Matters

This is a landmark enforcement action that sends several signals:

1. Conference policies have teeth — ICML is willing to reject hundreds of papers, not just issue warnings
2. Reciprocal accountability — the penalty falls on the violator's own paper, creating a strong incentive for honest reviewing
3. Detection is improving — the conference developed its own detection methodology rather than relying on unreliable AI detectors
4. The scale is alarming — 506 reviewers breaking rules they chose is a significant percentage, suggesting either carelessness or systematic non-compliance

The Broader Debate

The ICML post acknowledges that the AI community is divided on LLM use in peer review. Issues include whether authors consent to AI-assisted review, whether LLMs improve or degrade review quality, and how to enforce policies at scale.

This enforcement action adds empirical weight to the argument that voluntary compliance alone is insufficient for AI usage policies in academic settings.

Source: ICML Blog | HN Discussion