AI · 1h ago
AI models cite retracted papers, can't detect post-training retractions
A developer tested 12 AI models on scientific citations and found that while they flag 82% of famous old retractions, they miss 100% of retractions that occurred after their training cutoff. The models cited retracted 2025-2026 papers as valid evidence. The author built an open-source tool called sourcecheck that resolves citations against registries to catch these failures.
Meridian48 take
The finding underscores a structural limitation of static models: no amount of scaling can make them aware of post-training updates, making a verification layer essential for any citation-critical use case.
Read the full reporting
The best AI models cite retracted papers, and they cannot know it →
DEV Community
ai-citationsretracted-papers