Evaluating large language models for assessment of psychosis risk.

Models inferred clinical high risk for psychosis status and estimated severity and frequency across 15 symptom domains, benchmarked against researcher-rated scores.

LLM-generated symptom scores showed good correlations with researcher-rated scores (ICC sev = 0.74, ICC freq = 0.75).

Generated summaries were largely faithful to source transcripts, with low rates of clinically relevant confabulation (3%).

While accuracy scaled with model size, smaller models achieved competitive performance with substantially lower computational cost.