Comprehensive Summary
Forensic psychiatry is an interdisciplinary subspecialty that blends biological, psychosocial, and legal perspectives to assess how psychiatric disorders affect legal capacities using interviews, history, and psychometrics. This study evaluates whether a large language model (LLM) like GPT-4o can accurately analyze forensic psychiatry experts’ reports by evaluating variables tied to criminal responsibility, social dangerousness, and competence to stand trial, testing LLM’s role in demonstrating transparency in clinical narratives for decision support. The researchers applied a fixed set of queries (defendant history, mental status, psychiatric diagnosis, and legal decision making) to a Retrieval-Augmented Generation (RAG) pipeline that chunked and generated structured outputs to two Italian case reports. Case 1’s expert diagnosis of intellectual developmental disorder, antisocial personality disorder, and conclusions of substantially diminished responsibility and incompetence to stand trial closely aligned with LLM’s outputs of the diagnosis and most importantly the supporting clinical factors of impaired judgment, intellectual disability, and a history of aggressive and violent behavior. Case 2’s expert diagnosis of adjustment disorder with persistent mixed disturbances and conclusions of full criminal responsibility and competence to stand trial was correctly reflected by the LLM, highlighting the defendant’s full understanding of prosecution and trial dynamics, and lack of psychotic comorbidities, ultimately declaring the inability to constitute for medical-legal disability. Overall, it was found that the system consistently recovered the key elements and summarized them coherently, but struggled to determine which elements were most important for legal implications without prompt engineering. The potential for LLMs in forensic psychiatry is promising, but concerns about privacy, bias, transparency (black box risk) of LLMs, and occasional “hallucinations” or differing “temperature” of AI point toward a need for clearer guidelines for retrieval and prompt guidelines.
Outcomes and Implications
The potential implementation of LLMs in forensic psychiatry is worth serious study due to the lack of reproducible and transparent workflows, despite the fact that the field shapes sentencing, security measures, and clinical management. LLM text extraction tools can help assist in standardization of documentation and enable comparisons of population based research on how diagnoses, risk factors, and legal dispositions relate. In practice, an LLM pipeline could function as decision support while never replacing expert judgment. An LLM pipeline can flag inconsistent elements in documents, extract core data to standardize reports, and provide transparent reasoning chains for legal decision making. The RAG-based approach can mitigate the “black box” effect, where AI systems produce results without any clear explainability. Any courtroom-adjacent use will require cautious validation, bias mitigation, explainability, and ethical safeguards to ensure consistency, fairness, and reliability.