By Connie · This article contains affiliate links. We may earn a commission at no extra cost to you if you sign up through our links.

ICLR 2026 Rejects 497 Papers for AI Policy Violations — and 21% of Peer Reviews Were AI-Generated

Academic publishing has an AI problem. And the top conference in the field that built the AI is now the clearest proof of it.

The International Conference on Learning Representations (ICLR) 2026 received a record-breaking 20,000 submissions. It also became the first major conference to publicly confront what happens when the researchers who build AI start using AI to shortcut the process of publishing about it.

What Happened

ICLR 2026 introduced its strictest-ever AI disclosure policy: authors must declare all LLM use in both submissions and peer reviews. Any paper that uses AI extensively without disclosing it faces immediate desk rejection. Reviewers who use AI to write reviews without disclosure violate reviewer agreements.

The result: 497 papers — roughly 2% of all submissions — were desk-rejected before even entering review. Most violations involved using LLMs to draft paper text or rewrite results without disclosure.

The peer review side is even more alarming. An independent analysis by Pangram Labs found that 21% of all peer reviews submitted to ICLR 2026 were fully AI-generated. More than half showed some degree of AI assistance. This is not a fringe phenomenon — it is majority behavior on the reviewer side.

The Scale of the Problem

Why This Is Happening Now

The timing is not coincidental. AI models in 2025 and 2026 crossed the threshold where they can produce plausible academic prose indistinguishable from human writing. A researcher under deadline pressure faces a tempting shortcut: let the model write a first draft, clean it up, submit.

On the reviewer side, the math is even more compelling. Peer reviewers are typically unpaid volunteers reviewing 3–6 papers per conference cycle, each requiring hours of careful reading. When those reviewers are themselves under time pressure and facing 20,000 submissions to be parsed across thousands of volunteer slots, the temptation to offload the review to GPT or Claude is real.

The result is a collapse of the quality signal that peer review is supposed to provide. If 21% of reviews are AI-generated, those reviews are not assessing the work — they are pattern-matching against what a review is supposed to look like. Papers pass not because a human expert judged them sound, but because they triggered the right patterns in a language model that triggered the right patterns in another.

ICLR's Response and What Comes Next

ICLR 2026's approach is disclosure-first. The conference does not ban AI use outright — it bans undisclosed AI use. Authors who declare LLM assistance can proceed. Reviewers who disclose AI-assisted summaries are technically compliant.

This creates a strange incentive: the researchers most likely to be caught are the ones who didn't read the policy, not the ones who gamed it. A sophisticated actor can use AI, disclose it minimally, and pass. A naive actor who used Grammarly and forgot to mention it gets rejected.

Other major conferences are watching. NeurIPS 2026 is reportedly considering a stricter two-tier system: papers may use AI for editing but not content generation, with human co-authors required to certify the intellectual contribution. Whether that is enforceable is an open question.

The deeper question — whether peer review itself is still the right mechanism for validating AI research when AI is this capable — is being asked seriously for the first time.

How to Use AI Ethically in Research

Most conferences and journals now accept AI assistance — they require disclosure. Here is what is generally permitted versus what crosses the line:

See also: AI Scientist-v2: First AI Paper Published in Nature and How to Use AI for Research in 2026.