AI therapy: what the current evidence actually says

‘AI therapy’ is one of the fastest-rising mental-health searches, and the gap between the claims and the evidence is wide. Systematic reviews in JMIR find modest, real benefits for structured tasks — guided journalling, psychoeducation, light cognitive-behavioural exercises — and consistent warnings about everything beyond that.

Where it helps

The reliable use is reflection. A long-context chat thread is a patient, private place to write down how you actually felt this week and to be asked one good follow-up question. That is the Ledger layer of the AI Health Stack doing what it does best: turning scattered notes into a narrative you can re-read. It does not require a special app. It requires the discipline to write things down and the prompts to read them well.

Where it fails

The APA has been explicit that consumer chatbots can produce plausible but unsafe responses in moments of crisis, and that none should be presented as a replacement for a licensed professional. The model has no duty of care. This is exactly why every mental-health prompt in the method carries an explicit escalation clause: the AI drafts observations, a human decides, and anything resembling crisis goes to a qualified professional or emergency service — not a chat window.

Used inside those limits, AI is a genuinely useful instrument for understanding your own mind. Used as a substitute for care, it is a liability. The difference is literacy, not the tool.