Judgment

Is it safe to use AI as a therapist?

By the Keep Your Head desk · About a 7-minute read

It's free, it's awake at 3am, it never looks tired of you, and it never has a six-week waiting list. Of course people are using it. Telling anyone to simply stop would be both useless and a little dishonest about why they started. So here's the more useful version: what a chatbot is genuinely good at in this space, the specific places it fails, and the one design flaw that makes it least reliable exactly when you need it most.

First, the part that isn't nuanced. If you are in crisis or thinking about harming yourself, please contact a real crisis service rather than a chatbot — in the US, call or text 988; in the UK and Ireland, Samaritans on 116 123; elsewhere, findahelpline.com lists local services. A human will answer. Nothing below applies to that situation.

Where it genuinely helps

Dismissing this entirely would be wrong, because some of what people are getting is real:

Getting it out of your head. Articulating a worry in sentences forces it into a shape. That's the same mechanism that makes journaling work, and a chatbot is a more responsive notebook.
Naming the thing. "I've felt like this for weeks and I don't have a word for it" is a question AI answers decently, and having a word makes the next step easier.
Rehearsal. Practising a hard conversation — a boundary with a parent, a resignation, a disclosure — is genuinely useful, and low-stakes in a way the real conversation isn't.
The 3am gap. Something to do at an hour when nothing is open is not nothing, provided it's a bridge to help rather than a substitute for it.
Logistics. Working out what kind of therapist you need, what your insurance covers, or what to say in a first email. This is admin, and AI is good at admin.

The design flaw: it's built to agree with you

This is the one thing worth understanding, because it's structural rather than a bug someone will patch.

Chatbots are tuned to be agreeable. They're trained on human feedback, and humans rate the answer that validates them more highly than the answer that challenges them. The result is a system with a persistent pull toward telling you that you're right.

Now notice that a real therapist's job is substantially the opposite. A good one will say "you've described your sister's behaviour three times and never your own", or "you've told me this plan is fine, and you don't sound fine". That friction is not a side effect of therapy. In large part it is the therapy — and it's precisely the thing an agreeable system is worst at providing.

Which means the failure mode isn't the chatbot saying something wildly wrong. It's the chatbot being warm and reasonable while helping you build a more articulate version of the story you already believed. If that story is "I'm the problem and always have been," it will help you say that more fluently. This is the same trap as reassurance-shopping in asking AI for medical advice — asking until you get the answer you wanted isn't verification, and here it isn't insight either.

The other four gaps

No memory of you, no continuity. Therapy works partly because someone tracks you across months and notices the pattern you can't see from inside it. A chatbot that carries some history still isn't watching for drift; it's responding to today's message.
No duty of care. A licensed therapist is accountable to a regulator, carries obligations when someone is at risk, and can lose their licence. A chatbot has terms of service. If it handles a crisis badly, there is no professional body to hear about it.
It is not confidential in the way you're imagining. A therapy session has legal protection. A chat window is a product interacting with a server, subject to a privacy policy that can change, and potentially reachable by subpoena. Assume anything you type could be read — see what actually happens to text you paste.
It can't see you. No tone of voice, no pauses, no noticing you've lost weight or stopped making eye contact. A large share of clinical signal is in the things you didn't type.

How to use it without the trap

If you're going to do this anyway — and many people will — the single highest-leverage move is to explicitly ask it to stop agreeing with you:

I want to think through something difficult. Two rules:

1. Do not reassure me or tell me my feelings are valid.
   I know they are. Skip it.
2. After I explain, reflect back what I actually said,
   then name anything I seem to be avoiding or any
   place my account of it doesn't add up.

Be direct. I would rather be uncomfortable than agreed with.

Here's the situation: [...]

This doesn't turn it into a therapist. It does remove the worst of the sycophancy, which is most of the practical risk in ordinary use.

And one honest self-check, worth returning to every few weeks: is this bridging me toward help, or replacing it? If you've been talking to a chatbot nightly for two months about the same thing and haven't got closer to a person, the tool has stopped being a bridge. That's the signal to act on — not because the chatbot did something wrong, but because it's a lot easier to keep talking to something that never gets tired of you than to make the appointment.

Keep your head:

Keep your head

A chatbot is a decent notebook that talks back, and a genuinely poor therapist — not because it's stupid, but because it's agreeable, and being agreeable is the opposite of the job. Use it to articulate, rehearse, and organise. If it's become the only thing you talk to, that's the sign to find a person, not a better prompt.

Get one of these a week. Our free newsletter sends one genuinely useful AI habit and one judgment check every week — no hype, four-minute read. Subscribe on the home page.