Why estimates of hallucinated citations are probably low (veruscite.com)

Recent analysis suggests that previous estimates of AI-generated citation hallucinations in academic papers may be significantly understated. Research conducted by Zhao et al. (2026) reported varying rates of hallucinations across platforms like arXiv (0.39%) and PubMed Central (0.27%). However, the findings from Topaz et al. (2026) indicate a dramatic increase in fabricated references within biomedical literature, raising concerns in the academic community. A potential reason for the underestimates is that previous audits primarily focused on title discrepancies, missing more complex errors such as incorrect authorship and misattributed publication venues. In contrast, a new tool developed by VerusCite addresses these shortcomings by examining a broader spectrum of citation errors—considering author names, publication types beyond journal articles, and performing web searches when traditional databases yield no results. As shown in a small sample from recent publications flagged for hallucinations, nearly half of the errors involved real papers with incorrect author associations, suggesting that the prevalence of AI hallucinated citations could be much higher than previously thought. This research calls for further investigation, hinting that the true level of citation hallucinations may warrant a doubling of current estimates, posing significant implications for the integrity of academic research.