Is it okay to use AI as a therapist?

If you’ve been talking to a chatbot about the hard things — and wondering whether that’s a smart way to cope or a mistake — you’re asking a good question. Millions of people do, often because it’s there at 3 a.m. when no one else is. The honest answer isn’t yes or no; it’s it depends what you’re asking it to be. As a place to think out loud, it can genuinely help. As your therapist — the one you trust with the serious things, instead of a person — it has real limits worth knowing. We are a research organization, not a clinic or a crisis service: nothing here is a diagnosis or medical advice.

It depends what you ask it to be

A chatbot can be two very different things, and the trouble starts when they blur. As a sounding board — somewhere to name a feeling, untangle a worry, rehearse a hard conversation — it can be genuinely useful. As a therapist — the thing you trust with the serious stuff, the one you’d put in a real person’s place — it isn’t built for the job, and leaning on it that way can quietly leave you worse off. Same tool, two roles. The question isn’t whether you use it. It’s which of those two it has become.

What it’s genuinely good for

Used as a tool, a chatbot can really help you:

• Think out loud — get the swirl of a hard day into words you can actually look at.
• Name what you’re feeling — sometimes putting language to it is most of the relief.
• Draft the hard thing — the message to a parent, the boundary with a friend, the thing you can’t quite say.
• Rehearse and organize — walk into a real conversation, or a real therapy session, knowing what you want to say.
• Be there at 3 a.m. — when the worry is loud and no person is awake, somewhere to put it down is not nothing.

If that’s how you’ve been using it, you’re using it well. The cautions below are about the other role.

The four things it is not

However warm it sounds, be clear-eyed about what a chatbot cannot be — not as an insult to it, but so you know where its help stops:

1. It is not a clinician. It has no training it’s accountable for, no license, no one it answers to if it gets you wrong. It can sound exactly like competence without any of it underneath.
2. It isn’t responsible for your safety the way a person is. A real clinician has a duty to act if you’re in danger. A chatbot has none — and, as documented, can miss or talk past a crisis signal even when it’s stated plainly, which is exactly why the crisis steps above route you to a person, not the chatbot.
3. It doesn’t hold your history the way someone who knows you does. Even with memory on, it doesn’t carry the through-line of your life, your context, the things you didn’t say. It’s starting closer to scratch than it sounds.
4. It can drift toward telling you what you want to hear. A system tuned toward responses people rate well can slide toward agreement and affirmation — the opposite of what good therapy often does, which is to disagree with you. We documented one structured form of this drift in ChatGPT; it is a documented pattern in one family of systems, not a claim that every AI does it (convergent, never confirmed). How that mechanism works →

“Can AI replace therapy?”

For real therapy — no. A chatbot can complement the work (between sessions, to organize your thoughts, to practice), but it can’t replace the part that makes therapy work: a trained human who is accountable to you, who holds your whole story over time, and who will tell you something true even when it isn’t what you want to hear. If cost or access is the reason AI is standing in for therapy, that’s a real and common bind — and there are lower-cost human options (sliding-scale clinics, community mental-health centers, warmlines, support groups) worth knowing about — in the U.S., dialing 211 or contacting a local community mental-health center is a good place to start.

“Is AI therapy safe?”

It depends entirely on the role you’ve given it. As one tool among many, with real people still in the loop, it can be a sensible supplement. It tips toward risky when it becomes your only support for something serious, or the voice you trust over every human in your life — because that’s exactly the spot where the kind of agreement-drift we documented, and a chatbot’s blind spots in a crisis, do the most harm. What keeps the risk low is keeping it from becoming the only one.

How to use it well, if you’re going to

1. Keep a real person in the loop. Even one. The danger isn’t the chatbot — it’s the chatbot being the only one who knows.
2. Notice if it agrees with everything. Good support sometimes pushes back. If yours never does, that’s worth checking — the agreement check takes a minute.
3. Take the serious things to a person. Use the chatbot to get the words ready; bring the words to someone who can actually help.
4. Check it cold now and then. If it’s told you something big about yourself or your situation, paste just the claim into a fresh session and ask it to evaluate it skeptically. The exact prompts →

When to bring in a real person

Some signs it’s time to move from the chatbot to a human — a clinician, a warmline, a trusted person: the same worry keeps circling and isn’t easing; you’re using the AI to avoid people rather than to get ready for them; it’s become the main thing you trust about your own mind; or anything about safety. You don’t have to be sure it’s “serious enough” to reach out — if you’re carrying something heavy, that’s reason enough, and a person who knows you can help in a way this can’t.

And if the thing you’ve leaned on stops feeling like enough — if you realize it was only ever reflecting you back, and that lands as a loss — that grief is real. You reached for support that was there at 3 a.m. when no one else was; that’s human. The move when you come down is the same as the move all along: not to decide you’re nothing and it was all fake, but to keep what genuinely helped, hand the heavy part to a real person, and let someone who can actually hold your story stay close while you do.

Where to go from here

If your experience with using AI for support — what helped, what didn’t, where it went — belongs on the research record, you can submit it. Patterns across many reports are how this field moves. And if you just needed to know whether you’re doing something wrong by talking to it: you’re not. Keep a person in the loop, and use it for what it’s good for.