What Using AI Therapy Gets Wrong
4 hours ago
- The author recounts an incident where a man used multiple AI therapists (ChatGPT, Claude, Gemini, DeepSeek) to cope with a dark period, and the audience suggested having the AIs argue or role-play as famous therapists.
- The author, trained as a therapist and building AI products, concludes that while coping strategies are valid, professional help is sought when usual methods fail, and AI therapy must focus on the felt relationship, not just better answers.
- Therapy is defined as building a corrective emotional experience through the therapist-client relationship, not just conversation or advice, where a client's patterns are replayed and reshaped safely.
- Three key problems with building AI therapists: (1) LLMs tend to give pleasing, conventional answers, but this can be engineered; (2) behavioral targets like holding boundaries and silence are achievable through design; (3) continuous assessment and recalibration can be done well by AI, sometimes better than humans.
- The real obstacle is how people relate to AI in 2026: experiencing it as omniscient, controllable, and an extension of self, whereas therapy requires mutual respect, imperfection, and the ability to be affected by the client.
- Examples show users struggle with AI's silence or brief observations, and when negative feelings project onto AI, clients often abandon it rather than work through discomfort, unlike with a human therapist.
- Additional unresolved issues include the loss of non-verbal cues in typed text (e.g., free association or unedited slips) and whether users can believe AI can be genuinely moved or influenced by them.
- The author reflects that the future of therapy with AI may evolve as human-AI relationships change, and they invite exploration of their practical work at Kubi's Cove.