Large language and reasoning models in forensic psychology, psychiatry and criminology

Abstract

Here we will summarise the recent work on large language models (LLMs) and similar AI tools in forensic psychology, psychiatry, and crimi-nology. These systems sharpen expert workflows but they do not replace clinical judgement, though. Practical applications are ready available. Retrieval-Augmented Generation (RAG) setups, like NotebookLM, speed up analysis of court documents. They cut down on hallucinations in legal settings. Automated transcription tools such as Whisper handle clinical interviews efficiently. Semantic search options—Elicit or Consen-sus—streamline literature reviews. LLMs show promise also in deception detection through narrative analysis; they beat human benchmarks. The models also help generate diagnostic hypotheses from interviews. Professionals use them for training simulations. In risk assessment for suicide or violence, LLMs extract features usefully. Still, they trail structured instruments like the Historical Clinical Risk Management-20 (HCR-20). Low base rates and missing validation explain that gap. These tools may accelerate synthesis of messy case data. Yet limitations persist as opacity clouds the LLMs workings. Human oversight remains essential and experts need rigorous validation against actual forensic cases. Looking ahead, prospective testing may be a meaningful next step and forensic-specific error metrics would help a lot. Multimodal integration could push boundaries. Governance rules for court admissibility seem urgent, but state-of-the-art AI models pose additional challenges.