Fabricated citations are making it into published, peer-reviewed computer science papers — and nobody noticed until now.
Researchers built a verification pipeline called RefChecker and ran it against camera-ready accepted papers from ICLR, ICML, NeurIPS, and USENIX Security. They were looking for hard failures only: references to papers that do not exist, or that list substantially wrong authors. Even under that strict definition, roughly one in twenty NeurIPS and USENIX Security papers from 2025 contained at least two hallucinated citations. They also found a post-ChatGPT uptick across several venues, a tail of papers with five or more bad references in a single bibliography, and hallucinated citations in award-winning work.
The finding matters because citations are the connective tissue of science — they're how a field tracks what it knows and how it knows it. A hallucinated reference is not a typo; it points to a work that may not support the claim it's attached to, or does not exist at all. If this gets normalized, the entire literature becomes harder to trust.
The one semi-encouraging note: the authors report RefChecker costs about $0.04 per paper at venue scale, which means the fix is cheap enough to mandate — it just hasn't been yet.