Is AI therapy safe? An honest answer — and the questions to ask before you trust one.

“Is AI therapy safe?” is one of those questions where the instinct is right and the wording is wrong. The instinct — pause before you hand your hardest thoughts to a machine — is exactly the instinct to keep. The wording fails because there is no single thing called AI therapy. There is a branded therapy bot. There is a general chat tool you are using after midnight. And there is a human practitioner who happens to use software. Those three are not variations on a theme; they fail in completely different ways, and lumping them together is how people end up either too trusting or too dismissive.

So this is the honest version. Not “AI bad, humans good,” and not the launch-day promise that a chatbot has cracked mental health. Where an AI conversation is reasonably safe, where it is quietly not, which AI is actually best for the job, and the four questions to ask before you trust any of them.

first, separate the three things people mean

Before you can answer “is it safe,” you have to say which it. Each carries a different risk, and a different person on the hook when something goes wrong.

1. A general chat tool used for support. The 3 a.m. listener. Cheap, available, no clinical guardrails, no duty of care, and no idea it is being used this way. Safe for naming, structuring and drafting; unsafe the moment you treat its fluency as judgement.
2. A branded “AI therapy” app. A product wrapping a model in therapy language. Sometimes built with clinicians, often not. The marketing implies care; the terms of service usually disclaim it. The risk here is the gap between what it sounds like and what it legally promises — read that gap carefully.
3. A human practitioner using AI. A trained clinician who uses tools for notes, drafting or research, while holding the assessment, the risk, and the responsibility themselves. This is the only one of the three where someone qualified is accountable for the outcome.

Most of the fear in “is AI therapy safe” is aimed at the first two while quietly imagining the safety of the third. Keep them separate and the question gets answerable.

where an AI conversation is reasonably safe

Used as a thinking aid rather than a clinician, a general chat tool does three things well, and none of them is therapy.

• Naming — handing you the word for what is actually going on. “I am not sad, I am bracing.” The word is usually most of the early relief, and a patient thread is unusually good at finding it.
• Structuring — running a known structure faithfully. A CBT thought record, an ACT defusion prompt, a set of Socratic questions. Structures travel well into chat; the model can hold one at 3 a.m. without getting tired.
• Drafting — externalising the email you cannot send or the boundary you cannot phrase. A draft is not a decision. Seeing the shape of the thought, then deciding in the morning, is a low-cost, low-risk use.

where it quietly is not safe

The unsafe part is rarely the obvious howler. Models are fluent, agreeable and infinitely patient — and those exact qualities are where the harm lives. The risks compound slowly, which is why they are easy to miss.

• It normalises rumination. A human eventually changes the subject. A model will circle the same distress with you for an hour, mistaking repetition for support and giving you no new insight.
• It colludes with avoidance. Ask it to help you not have the hard conversation, and it will help you not have the hard conversation — warmly, and at length.
• It produces confident, mistaken framing. Plausible medical or psychological language with no mechanism for knowing it is wrong, delivered in the same calm tone as the parts it gets right.
• It cannot manage acute risk. Suicidality, abuse, a medical emergency — a chat thread can recognise keywords at best. It cannot assess, escalate, or take responsibility, and it must never be the thing standing between you and help.

so which AI is actually best for the job?

“Which AI is best for therapy” is the most common follow-up, and the honest answer disappoints people who want a brand name. The best tool depends on which of the three jobs you are doing — and none of them is a therapist. We frame it the way we frame the rest of the AI Health Stack: by the job, not the logo.

1. For understanding the evidence — what a technique is, what the research actually says, what the risks are — use a research-grade model and ask it for citations you can check, not reassurance.
2. For structuring and drafting — running a thought record, drafting the email, keeping a private log of mood and sleep — a general assistant is fine, and the one you already pay for is usually the right one. Switching brands rarely changes the outcome here.
3. For anything clinical — assessment, formulation, a plan that changes as you do — the best tool is a human practitioner. If you want AI in that loop, the right configuration is a clinician using AI, not an app standing in for one.

Notice what that does to the question. The “best AI for therapy” is not a better chatbot — it is a clearer division of labour between the tools you already have and the human you cannot replace.

the four questions to ask before you trust any of them

Whichever tool you are looking at — a chat thread, a branded app, or your own setup — run it through these four before you lean on it. They are the same questions a careful practitioner would ask on your behalf.

1. Who is accountable if this is wrong? If the answer is “no one” or “you accepted the terms,” treat it as a thinking aid, never as care.
2. What happens to what I type? Sensitive mental-health text is the most personal data you own. Check whether it is stored, used for training, or exportable — and prefer tools that let you keep and delete your own record.
3. What is it good at, and does this task match? Naming, structuring, drafting: safe. Diagnosing, treating, holding risk: not. If the task is in the second group, the tool is the wrong one regardless of brand.
4. When does it tell me to stop and find a human? A tool that never points you outward is a tool that will keep you in the loop forever. The good use of AI here ends with a hand-off, not a subscription.

the calm conclusion

Is AI therapy safe? As a thinking aid, on an ordinary hard night, used for naming, structuring and drafting, with your data in your control and a human in the loop for anything clinical — yes, reasonably. As a replacement for assessment, risk-holding and the relationship that does the actual healing — no, and no amount of fluency changes that.

The instinct to ask the question was right. Keep it. Then do the more useful thing: stop looking for the one safe app, and start using the tools you already have for what they are good at — and a human for what they are not.

“A model can listen forever. Listening forever is not the same as helping you change.”

— — the working rule we give readers