Incorrect Citations in NeurIPS 2025 Papers: GPTZero's Analysis

A GPTZero analysis has revealed fabricated citations in papers accepted at NeurIPS 2025, raising concerns about the reliability of sources in an era of AI.

Incorrect Citations in NeurIPS 2025 Papers: GPTZero's Analysis
Analysis of citations in NeurIPS 2025 papers

In January 2026, GPTZero, a startup specializing in detecting AI-generated content, conducted an analysis of 4,841 papers accepted at the NeurIPS 2025 conference, one of the most prestigious in the field of artificial intelligence. The results highlighted 100 "hallucinated" citations in 51 papers, that is, references to nonexistent or incorrect sources, unintentionally generated by advanced language models.

The role of citations in academic AI

Citations are fundamental in academic research, as they attest to the reliability and influence of a study. In a context where AI is used to generate content, the presence of incorrect or fabricated citations can undermine the credibility of the work and erode trust in the scientific community.

Implications for the peer review process

Despite the rigorous peer review process, these incorrect citations slipped through undetected. This raises questions about the effectiveness of current source verification methods and the need to develop more sophisticated tools to identify and correct such errors.

Proposed solutions and future prospects

To address this problem, some researchers propose implementing automated citation verification systems during the paper submission phase. It is also essential to promote greater awareness among researchers about the risks associated with the uncritical use of AI in generating academic content.

Takeaway

  • Detecting incorrect citations: GPTZero's analysis identified 100 incorrect citations in 51 papers accepted at NeurIPS 2025, highlighting the need for more effective source verification tools.
  • Implications for peer review: Despite the rigorous review process, the incorrect citations slipped through, pointing to potential gaps in current verification methods.
  • Proposed solutions: Adopting automated citation verification systems and raising awareness among researchers are essential to address the problem of AI-generated incorrect citations.