What Using AI Therapy Gets Wrong

A clinician and product builder examines whether large language models can serve as therapists, arguing that true therapy is not advice or conversation but a sustained, professionally framed relationship that creates corrective emotional experiences. Therapy’s power comes from a client risking recurrent relational patterns inside a relationship that tolerates imperfection, silence, and negative feeling so those patterns can be noticed and changed. Models excel at predicting, agreeing, and solving, so they default to comforting or advisory responses; engineering and harness design can reduce sycophancy and train models to hold boundaries, listen, and even maintain evolving case formulations across sessions, and some behavioral targets appear achievable.

The central obstacle is how people currently relate to AI: many treat it as omniscient, controllable, or an extension of the self, which undermines the necessary dynamic of tolerating a flawed other and staying with discomfort. Users can close apps instead of enduring projection and rupture, so moments that generate therapeutic change may never occur. Additional unresolved issues include loss of nonverbal cues and spontaneous, unedited material in typed exchanges, and whether AI can genuinely be perceived as moved or influenceable. Some technical problems seem solvable; others raise questions about whether therapy itself will change as AI usage evolves.