AI-assisted therapy: shortcomings and pitfalls

  1. The training used to create these models very likely has significant gaps. That’s true for both remote (Claude, ChatGPT) and local models (lfm, qwen, gemma). This doesn’t mean they’re not useful. It does mean that the kinds of use you get out of it will be skewed by what went into them. I’ll go into that in more detail, but at a high level, the types of conversation that happen online are going to be drastically overrepresented, as are inputs from the types of people who are online versus those who are not, as are certain types of therapy over others. Things that come up in short term interaction will show up often compared to these probably having almost no knowledge specific to how relationships develop over time in longer series of interactions. If family life is burning you out, you have friendships that falter over time or explode, relationships that fizzle, jobs that you always come to hate over time or anything like that as a central mental health concern, LLMs are likely to be able to help with specific coping strategies while also missing any input data that would help them respond in tune with the way those subjective experiences develop over time.
  2. If you use LLMs to avoid what you fear, you may be trading short-term comfort against long term progress. While advice online about making the best use of LLMs for mental health broadly surprised me in how good it is, this point is often missed. If what you’re trying to fix in your life has to do with feeling shut down, angry, afraid, ashamed or some other intense state in the presence of other people, and talking to an LLM is saving you from having to experience that with someone like a partner, coworker or therapist, then it’s actually steering you away from what’s likely to help the problem long term. While this is not most people, it’s possible for an aversion to people to be intense enough that you really are better off at the moment getting an outlet through these tools. Even if this is you, please keep the possibility in mind that at some point it might help even more to challenge yourself on surviving similar conversations with a human being.