Can LLMs actually conduct Socratic questioning in therapy?

For certain use cases, LLMs show promising ability to conduct individual tasks needed for psychotherapy — assessment, psychoeducation, demonstrating interventions. But clinical products and prototypes have not demonstrated anywhere near the sophistication required to take the place of psychotherapy. The gap is specific: while an LLM can generate an alternative belief in the style of CBT, it remains unproven whether it can engage in the turn-based, Socratic questioning that would be expected to produce cognitive change.

This distinction — between exhibiting a skill and implementing it therapeutically — is the core challenge. Generating an alternative belief is a single-turn text generation task. Socratic questioning requires tracking the patient's cognitive state across turns, calibrating the timing and intensity of challenges, adapting when the patient resists or deflects, and knowing when to push versus when to support. This is a multi-turn planning problem with a moving target (the patient's evolving understanding).

Since Can language models understand without actually executing correctly?, the therapy skill gap may be an instance of this broader pattern. LLMs can comprehend what Socratic questioning looks like (they can describe it, generate examples) but cannot competently execute it in live interaction. Psychotherapy transcripts are likely poorly represented in training data, and privacy/ethical concerns make such representation challenging. Prompt engineering may be the most feasible approach, but it cannot substitute for the adaptive multi-turn reasoning that therapy demands.

The five major challenges for LLM mental health deployment — hallucination, interpretability, bias, privacy, and clinical methodology — compound this gap. Intrinsic hallucination (contradicting dialogue history) directly undermines the internal consistency essential to therapeutic trust. The inability to process nonverbal cues removes a critical information channel. And the tendency to be overly prescriptive clashes with current evidence-based practice, which favors exploratory over directive approaches.