Researchers build a better detector for when vision AI hallucinates — using the AI's own internal thinking

Researchers found that combining multiple internal signals from vision-language models detects false claims more reliably than single-detector approaches. This means systems using image and text together could catch their own mistakes before users see them.

Vision-language models — the systems that look at images and answer questions about them — generate confident-sounding lies with regularity. Detection methods that run after the model outputs its answer are slow and imprecise. This work suggests looking at what the model is actually computing internally (its attention patterns, hidden states) catches hallucinations with higher accuracy. The practical implication: if deployed, this could reduce the false claims these systems generate in production, though only if someone actually runs the detector on every output.