4 ways AI could support psychotherapy

University of Utah researchers led by Zac Imel, with coauthors Vivek Srikumar and Brent Kious, published a framework that defines four levels of automation for applying AI in psychotherapy and posted it in Current Directions in Psychological Science. The work reframes the question from Will robots replace therapists? to What exactly are we automating and how much? The authors place automation along a continuum from scripted interventions to fully autonomous AI agents that interact with clients. Technical details: The paper grounds the taxonomy in real task decomposition and uses the self-driving car analogy to explain different risk profiles for each level. The framework distinguishes categories such as scripted delivery, clinician-assistive tools, analytics platforms, and direct-client AI. It highlights concrete capabilities, including: automated note-taking and session transcription with structured summaries; LLMs powering conversational prompts, guided coping scripts, and triage flows; analytics that annotate therapist behavior and produce feedback for supervision; direct-to-client conversational agents that could deliver interventions autonomously. The authors discuss implementation variables practitioners must consider: data provenance, model transparency, evaluation metrics for therapeutic outcomes, and human-in-the-loop controls. Context and significance: This framework arrives as LLMs and voice-enabled systems enter healthcare workflows. By translating abstract risks into task-level classifications, the paper helps clinicians and product teams align safety controls, consent models, and validation strategies with the degree of automation. The self-driving car analogy clarifies why mid-level assistive systems may offer value while reducing risks tied to fully autonomous therapeutic agents. What to watch: Practitioners should expect rapid growth in tools that occupy the clinician-assist and analytics categories; open questions include clinical validation, regulatory classification, data governance, and standards for performance metrics. The line between assistive and autonomous systems will shape procurement, liability, and patient safety decisions going forward.