Benchmarking large language models against human experts in rehabilitation medicine: a multidimensional evaluation.

The study compared AI to 'textbook' expert answers, not the variable and often hurried plans of clinicians in routine practice. No real patients were involved; this was a simulation using written case summaries, so it cannot predict actual health outcomes or patient safety in the real world. The expert panel was small (6 people) and recruited from high-level Chinese institutions, which may bias the preference for integrated Traditional Chinese Medicine approaches. The evaluation was text-only. It did not test the AI's ability to interpret imaging or observe patient movement, which are critical for physical and occupational therapy.

Funded by The First Affiliated Hospital of University of South China.

Do not interpret the high AI scores as evidence that AI is safer or more effective than a human therapist. The AI won on 'standardized textbook' criteria like structure and evidence citation. It failed to demonstrate the clinical judgment to prioritize a patient's psychological stability or adapt goals to their specific life context, which are the skills that actually drive rehabilitation success.