The original title is 11 words: "The best AI models cite retracted papers, and they cannot know it"

Short summary

Frontier AI models like GPT-5.5 and Claude variants correctly cite real, authoritative papers that were retracted after their training cutoff, flagging 0% of post-cutoff retractions versus 82% for famous old ones. The author built sourcecheck, an open-source tool that resolves every citation against OpenAlex, Crossref, and Retraction Watch, catching every retraction regardless of model knowledge. The key insight: no amount of scaling fixes this gap—only a registry lookup layer can, making verification infrastructure essential for any production system citing scientific, medical, or legal literature.

• •Top models flagged 82% of old famous retractions but 0% of post-cutoff retractions
• •sourcecheck resolves citations against OpenAlex, Crossref, and Retraction Watch to catch what models cannot
• •Scaling does not close the retraction knowledge gap; only a verification lookup layer does

Generated with AI, which can make mistakes.