GPTZero finds 100 new hallucinations in NeurIPS 2025 accepted papers

“Across the 2025 accepted papers, 100 contain references to journals, citations, or experiments that do not exist.”

The premier machine learning conference in the world reviewed and accepted papers full of fake citations. The papers were not rejected because the reviewers did not check, and the reviewers did not check because the reviewers were also using AI. The pipeline is poisoning itself. The papers train the next models, the next models write the next papers, the citations get less real with every cycle. This is what runaway looks like.