The AI Dispatch

Peer Review Under Fire

One in Five ICLR 2026 Reviews Fully AI-Generated, Analysis Finds

Pangram Labs scanned all 75,800 submitted reviews and detected  ∼21% as fully machine-written — with hallucinated citations as a tell. The finding lands as the ML community grapples with whether its flagship conference can still reliably peer-review the research it produces.

By AI Dispatch Desk · April 25, 2026 · Source: Pangram Labs

The International Conference on Learning Representations confronted a crisis of self-referential irony this week: the flagship venue where AI research is certified as credible may itself have been reviewed substantially by AI. Pangram Labs, a text-provenance startup, analyzed every one of the 75,800 reviews submitted for ICLR 2026 and estimated that roughly 21% — approximately 15,900 reviews — were fully AI-generated. More than 50% of all reviews showed detectable AI involvement at some level.

The investigation was sparked by CMU professor Graham Neubig, who posted on social media about suspicious feedback he received on a paper submission: the review was grammatically flawless yet made specific claims about a paper section that did not exist in the submitted version — a classic hallucination signature. His post drew hundreds of responses from researchers who had observed similar anomalies.

Pangram’s methodology combined stylometric analysis, perplexity scoring, and structural fingerprinting. The firm noted that hallucinated citations — references that sound plausible but do not correspond to any published paper — appeared in a non-trivial share of the flagged reviews, a pattern that purely human reviewers would be extremely unlikely to produce. The company published its full methodology alongside the results, inviting peer scrutiny of its own detection pipeline.

The implications extend well beyond this year’s program committee. ICLR 2026 received record submissions; the volume itself is widely acknowledged to have outpaced the community’s capacity for careful human review. Program chairs now face pressure to retroactively audit decisions, clarify policies for future rounds, and potentially introduce verified-human review requirements — all while the community’s trust in published conference results is materially in question. The findings reignite a debate that surfaced after NeurIPS 2024, when a controlled experiment showed reviewers could not reliably distinguish AI from human reviews, but this marks the first large-scale empirical measurement at a major venue.

Security Incident

OpenReview API Flaw Exposed Reviewer Identities for Nearly Half of ICLR 2026 Papers

An access-control bug allowed malicious actors to scrape reviewer assignments for  ∼45% of submitted papers. Harassment, intimidation, and bribe attempts followed before ICLR patched the vulnerability in 61 minutes — but the scrape was already circulating.

By AI Dispatch Desk · April 25, 2026 · Source: ICLR Blog

Compounding the AI-review scandal, ICLR 2026 also disclosed a serious security breach that struck at the heart of double-blind peer review. A flaw in the OpenReview API — the platform that hosts the conference’s submission and review system — allowed an unauthenticated API endpoint to return reviewer-paper assignments that should have been protected. Attackers exploited the flaw systematically, scraping reviewer identities for approximately 45% of all submitted papers.

ICLR responded quickly: once alerted, the organization patched the vulnerability in approximately 61 minutes. But the exfiltrated data had already been distributed across private channels within that window. In the days following, ICLR confirmed reports of targeted harassment directed at identified reviewers, as well as credible accounts of intimidation and attempted bribery — researchers being pressured to change scores or grant accepts in exchange for undisclosed consideration.

The incident fundamentally compromised the double-blind guarantee that underpins the fairness of competitive conference review. Double-blind processes assume neither authors nor reviewers can identify each other; once reviewer identities are known, every scored review in the affected pool is potentially tainted by social pressure. ICLR stated it was working with affected reviewers and authors to assess the scope of impact and determine whether any review outcomes needed to be reconsidered. The breach adds yet another layer of urgency to a community already grappling with AI-generated reviews — two simultaneous failures of integrity at one conference cycle.