Top AI Models Cite Retracted Papers With No Way to Detect It, Study Finds

A developer tested twelve leading AI models — including GPT-5.5 and Claude variants — by asking them to cite scientific literature, then verified every citation against live registries. While models correctly flagged roughly 82% of well-known older retractions they had learned during training, they failed to flag post-cutoff retractions 100% of the time, citing invalidated papers as valid evidence. The core problem is structural: no amount of model scaling can give an AI knowledge of retractions that occurred after its training cutoff. To address this, the developer built an open-source tool called sourcecheck, which resolves citations against OpenAlex, Crossref, and Retraction Watch in real time. The findings highlight the need for an external verification layer in AI systems that rely on scientific citations.