Before you trust an AI with your mental health: seven questions to run first.

Almost nobody decides to trust an AI with their mental health. It is not a decision so much as a drift — it happens at midnight, mid-spiral, with a tool that is open, fluent, patient and endlessly agreeable. Those last three qualities feel like care. They are not. They are properties of a language model, and they are exactly the properties that make the drift dangerous, because nothing in the experience tells you when you have crossed from “useful thinking aid” into “relying on something that cannot catch you.”

This is not an argument against ever using AI for this. Used carefully it can genuinely help. It is a pre-flight: seven questions to run before you lean on a tool with something tender, so the choice is one you actually made. Run them once on whatever tool you are tempted by — a chat thread, a branded “wellness” app, your own setup — and you will know in two minutes which lane you are in.

1. who is accountable if this is wrong?

Start here, because it sorts everything else. If something the tool tells you turns out to be harmful, who is on the hook? A licensed practitioner carries a duty of care and can be held to it. A chatbot carries a terms-of-service page that disclaims responsibility the moment you tap accept. If the honest answer is “no one” or “me, because I agreed,” treat the tool as a thinking aid and never as care — no matter how caring it sounds.

2. what happens to what i type?

Mental-health text is the most revealing data you will ever produce. Before you type the real thing, know whether it is stored, whether it is used to train the model, and whether you can export and delete it. Prefer tools that let you keep your own record and remove it on demand. If you cannot find out what happens to the words, that is itself the answer.

3. is the job inside the tool’s lane?

AI is genuinely good at three jobs near mental health: naming what you feel, structuring a known technique, and drafting the thing you cannot say. It is dangerous at a fourth: assessing how much trouble you are in. Before you trust it, name the job out loud. If the job is naming, structuring or drafting, you are in the lane. If the job is assessment — “is this normal, am I okay, how bad is this” — you have left it.

4. is it agreeing with me, or helping me?

A model is built to be agreeable, which means it will validate a distorted thought as readily as a healthy one. Ask yourself whether the last few exchanges introduced anything new — a reframe, a question, a gentle contradiction — or whether the tool has simply been nodding along for twenty minutes. Endless agreement is not support. A good human eventually pushes back; a model rarely will unless you make it.

5. what is its plan for an emergency?

Decide this while calm, because you cannot decide it in crisis. A chat thread can, at best, recognise a few keywords. It cannot assess acute risk, cannot escalate, and cannot take responsibility for keeping you safe. Know your real escalation path — a crisis line, a person, emergency services — before you ever need it, and never let a chatbot be the thing standing between you and that path.

6. is it making me more capable, or more dependent?

The quiet failure mode is not a single bad night; it is a slow outsourcing of your own regulation. Ask, honestly, whether the tool is teaching you something you carry away — a technique, a clearer thought, a draft you can send — or whether you are simply returning to it because the loop feels soothing. Help leaves you more able to cope without it. Dependency leaves you needing it to cope at all.

7. could i walk away from it tomorrow?

The last question is the exit. If you wanted to stop using this tool tomorrow, could you — keep your record, lose nothing essential, and still have your real supports in place? A tool you can leave is a tool you can trust to be one input among many. A tool you cannot leave has already become more than that, and that is the moment to bring a human back into the loop.

the point of the pre-flight

None of this is about being afraid of the tools. It is about making the decision on purpose instead of by drift — keeping the instinct that made you pause before handing your hardest thoughts to a machine, and turning it into seven questions you can actually answer. Do that, and AI becomes what it is good at being: a calm, private, capable aid for the jobs it can do, with a clear line where the human takes over.