Comment by appstorelottery

This is a multi-session problem, or at least this is how I've solved it in the past (i.e. the user chat/report thread isn't the one driving the conversation, the sub-sessions/workflows are guiding the analysis).

p.s. there's some interesting literature on how in some circumstances (i.e. religious delusions), where actually validating the patient - basically confirming their experience can lead to rapid recovery (however this approach is not found in western culture). I do agree with you in terms of the risks of LLM's driving folks deeper into delusional beliefs & behaviors - it's very real.