Designing Clinical Large Language Models for Evidence-Based Psychological Practice - PsyARXiv Preprints
Stade, E. C., Stirman, S. W., Held, P., Schwartz, H. A., & Eichstaedt, J. C. (in press). Designing clinical psychological AI that reduces suffering: Challenges and technical considerations.
Abstract
Large language models (LLMs) offer significant potential to augment or overhaul aspects of psychological assessment and treatment. However, current LLM technologies have yet to demonstrate the capacity to effect meaningful and sustained clinical change. This is due, in part, to LLM systems and products being developed with insufficient integration of clinical science expertise. To address this gap, here we identify important components of evidence-based psychological practice for integration into clinical LLMs: 1) psychodiagnostic assessment, 2) longitudinal case conceptualization, 3) appropriately dosed intervention planning, 4) meaningful progress evaluation, 5) attention to treatment sustainment, 6) clinically appropriate style, and 7) integration of emerging understandings of mechanisms. We introduce a set of key technical questions and considerations for development and evaluation of clinical LLMs against these guidelines. Despite their potential, current LLMs face notable limitations in meeting these guidelines, including issues with memory, sycophancy, and prioritizing short-term helpfulness over long-term clinical targets. Designing effective, clinical-science-based LLM systems requires understanding and carefully balancing the abilities and limitations of LLMs.