An AI-text detector estimated about 21% of ICLR 2026 paper reviews were fully AI-written, and over half had some AI involvement.

This explores how much of the reviewing at ICLR 2026 was actually written by AI rather than by the people assigned to do it. The most direct estimate comes from Pangram Labs. It ran an AI-text detector over ICLR's public reviews and estimated that about 21% were fully AI-generated and more than half had some AI involvement How much AI content appears in peer review at ICLR?. These are a detector's predictions, not confirmed cases. Still, the picture is clear: covert AI involvement was common, not rare.

An earlier study gives a baseline, though not a perfect one. It looked at ICLR 2024, NeurIPS 2023 and other AI conferences and estimated that 6.5% to 16.9% of review text had been substantially modified by an LLM How much peer review text shows signs of LLM modification?. That study measured the whole pool of reviews statistically instead of flagging individual ones. Because the two studies counted in different ways, the numbers aren't a clean trend line. They point the same way, though. The earlier study also found who leaned on AI most: reviewers who were less confident, rushed or less engaged. So AI tends to fill in where human effort is thinnest.

The more surprising finding is what AI involvement did to the reviews. In Pangram's data, reviews with more AI text gave systematically higher scores How much AI content appears in peer review at ICLR?. AI wasn't just polishing what a human had already decided; it may have been making reviews more generous. A related problem makes this worse. AI reviewers show a 'hivemind' effect, agreeing with each other more than human reviewers do, and authors can raise AI review scores by about 0.45 points just by having a model rewrite their paper, without improving the science Can AI systems safely replace human peer reviewers?. If a fifth of reviews come from similar models, a paper is getting fewer independent opinions than its review count suggests.

ICLR 2026's organizers chose not to punish based on detector scores alone. Detector flags went to area chairs as one signal among several, and sanctions required concrete human-found evidence How do detection tools shape LLM use enforcement at ICLR?. The firm enforcement point was hallucinated references: papers with confirmed fabricated citations were rejected without full review How can conferences detect and handle LLM misuse in peer review?. Compare AAAI-26, which took the opposite approach. It openly added one labeled AI review to every fully reviewed paper, and survey respondents said they preferred those reviews on technical accuracy Did conference reviewers prefer AI reviews over human ones?.

So the honest answer is that roughly a fifth of ICLR 2026 reviews may have been written entirely by AI, and most involved it somehow. The bigger issue is that this happened without disclosure, so nobody could adjust for the inflated scores or the shared blind spots. Some argue that if AI speeds up how fast research is produced, review will have to be AI-assisted too, and the real choice is whether that happens openly and with someone accountable Can human review keep pace with AI-accelerated research generation?. Others say the fix needs authors, reviewers and venues to share responsibility, for example by letting authors rate how good a review was before they see the decision Can two-stage review and badges fix AI conference peer review?.