Researchers have released MedHal, a large synthetic dataset built specifically to catch AI-generated falsehoods in medical text.
Hallucination detection tools trained on general text often stumble in medicine, where the same generic-sounding error can mean the difference between a correct dosage and a dangerous one. MedHal tackles this by pulling from diverse medical text sources and covering both intrinsic hallucinations (claims that contradict the source) and extrinsic ones (claims that can't be verified against it at all). The researchers trained a baseline hallucination detection model on the dataset and report it outperforms general-purpose detection approaches on medical text. The dataset is large enough to support training, not just evaluation, which is the harder problem to solve.
This matters because most hallucination benchmarks are built for chatbots answering trivia, not for systems drafting clinical notes or summarizing patient histories. A model that scores well on generic hallucination tests can still confidently invent a drug interaction or misstate a lab result. Cheaper, automated detection also means fewer medical AI tools have to rely on expensive expert review to get evaluated at all, which could speed up how fast this research area moves.
A detection dataset does not fix a model's tendency to make things up. It just makes it cheaper to catch when it does.