Applicability of LLM capabilities to complex psychiatric diagnosis

Determine whether large language models that achieve high performance on standardized medical benchmarks can be effectively applied to diagnosing complex psychiatric conditions that lack well-established objective biomarkers and rely on nuanced, subjective first-person accounts.

Background

Recent work shows LLMs can match or surpass clinicians on standardized medical benchmarks and assist in decision support, yet these tasks typically involve structured formats and clearer ground truth. Psychiatric diagnosis, by contrast, often depends on narrative reports and lacks objective markers.

The paper targets personality disorders (Borderline and Narcissistic) using first-person autobiographical narratives in Polish. Determining whether LLM capabilities transfer to this less structured, subjective domain is clinically significant given the growing public use of LLMs for self-assessment and the risks of misdiagnosis.