GPTZero finds 100 new hallucinations in NeurIPS 2025 accepted papers (gptzero.me)

Article summary

GPTZero's analysis of 4841 papers accepted by NeurIPS 2025 found at least 100 papers with confirmed hallucinations, which are fabricated or inaccurate citations. The discovery highlights a critical vulnerability in the peer review pipeline, where reviewers are outnumbered and outgunned by the volume of submissions. The issue is exacerbated by the use of generative AI tools, which can produce plausible but false papers. This problem can have significant consequences for the validity and reliability of scientific research.

Main themes

• Academic integrity
• Peer review pipeline
• Generative AI
• Scientific research validity
• Reproducibility crisis

What commenters say

• The proliferation of fake or inaccurate papers due to generative AI will harm scientific research and its credibility.
• Requiring reproducibility of research findings could help address the issue, but it is a challenging and resource-intensive process.
• Incentivizing reproducibility, such as through a separate track for reproduction papers, could be an effective solution, but it may also lead to bad reproductions and rejection of true theories.
• The current system prioritizes novelty over reproducibility, leading to a lack of validation of existing research and a focus on publishing new findings rather than verifying existing ones.
• Some argue that the problem lies not with reproducibility, but with the quality of the original research, and that requiring better research in the first place would be more effective.
• Others propose alternative solutions, such as a new degree or certification that focuses on validating and analyzing the state of the art in a field, rather than just producing new research.
• The issue of reproducibility is complex and multifaceted, and there is no easy solution, but it is essential to address it to maintain the integrity of scientific research.
• Some researchers believe that the current system is not broken and that citations can serve as a good proxy for validation, but others argue that this is not enough to ensure the reliability of scientific findings.