An analysis of nearly 76,000 peer reviews for next year’s International Conference on Learning Representations found that about 21% were fully generated by AI, with more than half showing some signs of AI assistance, according to detection firm Pangram Labs. The scan, covering roughly 19,500 submissions, also flagged 1% of manuscripts as fully AI-written and 9% as more than half AI-generated. The revelations, which followed mounting complaints from researchers about off-base critiques and hallucinated citations, have prompted ICLR organizers to deploy automated checks for policy breaches. Conference leaders say the move aims to shore up trust in a process under strain, as the use of large language models by reviewers becomes more prevalent.
Related articles:
ChatGPT listed as author on research papers — many scientists disapprove
DetectGPT: Zero-Shot Machine-Generated Text Detection
