Academic audits likely underestimate AI-hallucinated citations in preprints because they only check titles in online databases, missing errors in author attribution, journal placement, and non-journal sources. Analysis of 46 recent hallucinations shows nearly half involve real papers with wrong authors or journals, suggesting current prevalence estimates may need to be doubled.
Spanda is a Rust-based tool that detects LLM hallucinations and quantifies epistemic uncertainty in sub-microsecond time without requiring expensive neural cross-encoders. It uses Exact-Match Normalized Entropy to match or exceed traditional Semantic Entropy methods while operating ~90,000× faster, making it practical for high-throughput production serving.
A VerusCite tool developer criticizes Springer Nature's AI citation-checking system, arguing it fails to catch hallucinated references in published articles from August–September 2026. The author provides examples across multiple major publishers and argues that for $2 per paper, publishers should invest more in reference verification rather than relying solely on peer review.