Top AI Conference Overlooked 100+ Fake Citations in Peer Review Process
Recent findings from an analysis of nearly 5,000 research papers presented at the NeurIPS 2025 conference reveal a troubling trend in the peer review process of academic work. The prestigious Conference on Neural Information Processing Systems (NeurIPS) accepted a total of 5,290 papers, but a significant number contained fabricated citations, raising questions about the effectiveness of the peer review system.
Key Findings from the NeurIPS 2025 Analysis
The AI detection firm GPTZero examined 4,841 accepted papers and identified over 100 instances of fabricated citations across 51 different papers. This analysis underscores the vulnerability of even carefully reviewed academic submissions.
- Conference: NeurIPS 2025
- Total Accepted Papers: 5,290
- Reviewed Papers Analyzed: 4,841
- Fabricated Citations Found: Over 100
- Acceptance Rate: 24.52%
- Total Submissions: 15,000+
Examples of Fabricated Citations
The analysis revealed various forms of citation errors. One paper, titled “SimWorld,” included fictitious author names such as “John Doe” and “Jane Smith.” In another case, the paper “Unmasking Puppeteers” provided DOIs and URLs that led to nonexistent sources. Notably, one submission concerning semantic uncertainty had 15 fabricated citations, including incomplete arXiv IDs.
The Phenomenon of “Vibe Citing”
GPTZero has classified these misleading citations as “Vibe Citing.” This term refers to citations that appear valid upon initial review but collapse under closer investigation. Patterns of Vibe Citing include:
- Combining real and fictitious authors.
- Inventing URLs and DOIs.
- Altering existing titles slightly to create false citations.
Underlying Issues in the Peer Review Process
This issue of fabricated citations is compounded by a significant increase in submissions to NeurIPS, with numbers rising over 220% from 2020 to 2025. Organizers have struggled to manage this influx, leading to concerns about the quality of peer reviews.
Additionally, distrust among authors towards reviewers is evident. Some authors, such as those behind the paper “Efficient Fine-Tuning of Quantized Models,” have withdrawn submissions, claiming reviewers relied on AI tools instead of thoroughly reading their work.
Broader Implications of Citation Fabrication
The prevalence of fabricated citations is part of a larger systemic issue in academic research. A 2024 study published in Research Ethics explored toxic incentive structures in elite universities, emphasizing how pressure to meet bureaucratic goals can lead to ethical lapses among researchers. This pressure often results in a detachment from ethical standards, making it more tempting for researchers to engage in misconduct.
Reports indicate that institutions often overlook such misconduct as long as research outputs meet expectations. The findings from GPTZero illuminate the pressing need for improved oversight and accountability within academic publishing.
Future Investigations and Solutions
As concerns about the integrity of academic citations grow, GPTZero plans to investigate the appendices of research papers further. These sections are typically less scrutinized and may contain additional inaccuracies. Such investigations require rigorous fact-checking and validation of sources, surpassing mere statistical detection of patterns.
The issues that arise from fabricated citations serve as a reminder that maintaining the credibility of academic research must be a priority. Enhanced scrutiny and reform in peer review processes will be essential to ensure that future conferences uphold the integrity of scientific communication.
