- The ICLR 2026 (International Conference on Learning Representations)—the world’s premier machine learning conference—is in turmoil after it was discovered that 21% of the peer reviews submitted were written entirely by generative AI.
- The investigation was conducted by Pangram Labs (New York) following a call from Professor Graham Neubig (Carnegie Mellon University), who suspected that many of the reviews he received showed signs of “hallucinated citations” and were rambling, meaningless responses.
- Pangram scanned 19,490 research papers and 75,800 reviews submitted to ICLR 2026 using its AI detection tool. The results:
- 15,899 reviews (≈21%) were entirely generated by AI.
- Over 50% of the remaining reviews showed signs of AI-assisted editing.
- 199 research papers (≈1%) were suspected of being fully written by AI.
- 9% of the papers had more than half of their content generated by AI.
- Several researchers, such as Desmond Elliott (University of Copenhagen), confirmed that the reviews they received contained incorrect data, bizarre phrasing, and missed the central focus of the research—all hallmarks of responses generated by Large Language Models (LLMs).
- The ICLR organizing committee, led by Bharath Hariharan (Cornell University), stated that they will implement automated tools to verify the misuse of AI in reviews and submissions.
- While ICLR allows the use of AI for language editing and writing experimental code, it strictly prohibits its use for drafting academic content or disclosing confidential information.
- Some authors have withdrawn their submissions due to receiving misleading reviews, while many others are confused about how to respond.
- Neubig acknowledged that the review system is “overwhelmed” due to the submission volume increasing severalfold over the past five years; each reviewer must read five papers in two weeks, making reliance on AI an “easy way out” or “shortcut.”
- Pangram warns that its detection tool still has a rate of false positives, so handling violations requires caution and transparency.
📌 Summary
ICLR 2026—the world’s premier machine learning conference—is facing a shock as 21% of its peer reviews, equivalent to nearly 16,000 reviews, were found to be written entirely by generative AI. The received reviews contained incorrect data, bizarre phrasing, and missed the central focus of the research, which are characteristic features of feedback generated by generative AI.
