Designing Clinical Psychological AI that Reduces Suffering: Challenges and Technical Considerations

Artificial intelligence (AI), particularly large language models (LLMs), has garnered significant attention for their potential to augment and possibly even replace current forms of psychological assessment or treatment. However, current AI technologies have yet to demonstrate the capacity to effect lasting psychological change. This paper outlines five key criteria for developing effective AI clinical psychology interventions, grounded in clinical science principles and emerging evidence from the AI mental health field: 1) attending to mechanisms of patient change, 2) conducting meaningful evaluation, 3) attending to time and dose, 4) modeling key characteristics of therapists and patients; and 5) attending to engagement and sustainability. Despite their potential, LLMs face notable technical and therapeutic limitations, including memory constraints, failure to guide users through a coherent set of interventions, the tendency to engage in advice-giving over inquiry, and a proclivity to demonstrate rather than teach skills. We provide key technical considerations relevant to each criterion, including using multi-agentic models, using multi-step language programming, and finding more suitable objectives for reinforcement learning. Lastly, we discuss open questions in the field of clinical psychology AI, including whether humans will interact with AI the same way they interact with other humans, how scalability can be balanced with human monitoring, given safety concerns, and how large language models might radically disrupt how psychological interventions are delivered. Only by combining insights from clinical science with a clear understanding of the technical affordances and limitations of large language models will it be possible for AI to make a meaningful impact on human psychological suffering.