Fake AI Citations Found in Leading AI Conference Papers, Raising Research Integrity Concerns
Over 50 research papers accepted at NeurIPS were found to contain AI-generated fake citations, raising serious concerns about AI hallucinations, research integrity, and the growing misuse of AI tools.
Tech Author
Concerns over AI-generated content in scientific research have risen when dozens of research papers submitted at NeurIPS were discovered to include fake AI-generated citations. According to a new report, more than 51 papers contained references that did not exist, raising major concerns about research quality and academic credibility.
GPTZero shared the data after scanning 4,841 papers accepted at NeurIPS 2025, which was hosted in San Diego, USA. Among the 51 papers, the organization uncovered over 100 hallucinated citations (fake or non-existent references created by AI tools that look real but cannot be found in actual books, papers, or online sources). While the quantity may appear minor in comparison to overall submissions, NeurIPS regulations clearly specify that even a single fake citation is sufficient to reject or remove a paper.
NeurIPS is one of the world's leading conferences in artificial intelligence and machine learning. Having a paper accepted there is regarded a significant accomplishment, especially given the acceptance rate for main-track submissions was only 24.52%.This means that articles with fake citations outperformed hundreds of legitimate submissions, revealing an emerging issue with AI-assisted research.
GPTZero explained that many researchers now employ large language models (LLMs) to assist with paper writing. However, these techniques can occasionally generate references that sound genuine but do not exist. The company refers to fake references as "vibe citations", citations that appear credible but cannot be discovered online.
To discover these vulnerabilities, GPTZero deployed an in-house tool named Hallucination Check, which searches used sources and highlights unverified references. Each highlighted citation was thoroughly checked by humans to ensure accuracy. The company has now made this tool available to authors and editors to help them identify problems before publishing. This issue is not exclusive to NeurIPS. GPTZero also discovered over 50 fake citations in articles being reviewed for ICLR 2026.
Experts warn that while AI tools can improve productivity, they also increase the risk of errors and fake information. As AI becomes more common in research, stronger checks, better review systems, and responsible AI use will be essential to protect the future of scientific publishing.
This article is based on information from The Indian Express
