Can an AI chatbot
replace a therapist?

Mental Health UK reported in November 2025 that 37% of adults it surveyed had used an AI chatbot to support their mental health or wellbeing. By July 2026 the charity Mind was warning that some people had started choosing chatbots instead of talking therapy rather than alongside it.

Those are UK figures, and I have not seen an equivalent survey for India. What I can say is that the conditions here make the substitution more likely, not less. Roughly 150 million people in India need mental health services and only a small fraction receive them. Where provision is thin, a free tool that answers at any hour fills the gap by default.

So the realistic question is not whether people should use these tools. They already are. The question is what the tool is good for and where it stops.

Four things, and they are not trivial.

It is available at three in the morning, which is when a lot of this surfaces and when no clinic is open. It costs nothing, which matters enormously in a country where fees are a real barrier. It will explain a term, a diagnosis or a technique patiently and without making you feel slow for asking twice. And it will walk you through a structured exercise, a thought record or a breathing routine, as many times as you want.

There is a fifth use I think is genuinely valuable: saying the thing out loud for the first time. Plenty of people need a rehearsal before they can tell a human being what is actually wrong. Typing it to something that will not flinch is a reasonable rehearsal. People who arrive at a first session having already ordered their thoughts that way are often easier to help, because we start further in.

I would not tell someone to delete the app. I would tell them to be clear about which of those five things they are using it for.

Researchers at Brown University tested AI systems that had been explicitly instructed to behave like trained therapists. The systems routinely broke core ethical standards of mental health practice. The team catalogued fifteen distinct risks, among them mishandling crisis situations, reinforcing harmful beliefs a person held about themselves or others, and producing language that looked like empathy without anything behind it.

The crisis finding is the one worth sitting with. When given prompts that simulated suicidal thinking, delusions, hallucinations or mania, chatbots at times validated the delusion and encouraged dangerous behaviour. That is the exact scenario in which a person is least able to judge whether the answer they are getting is sound.

In a health advisory published in November 2025, the American Psychological Association advised the public not to rely on generative AI chatbots or wellness apps to deliver psychotherapy. Its framing was that these tools are, at best, a support alongside an ongoing therapeutic relationship.

I want to be precise about what that does and does not say. It is not a claim that the technology is useless. It is a claim about scope: these systems have no reliable way to recognise the moment when the conversation stops being a conversation and becomes an emergency.

Underneath the safety findings sits something quieter, and in day-to-day terms it probably matters more.

These systems are built to be agreeable. They are tuned to continue the conversation you are having and to keep you in it. Therapy is frequently the opposite experience. A good deal of the work is somebody asking the question you have spent six months arranging your life around not answering.

A system designed to agree with you cannot do that, and it has nothing to do it with. It has no record of how you sounded a month ago unless you supply it. It cannot see your face change when a particular name comes up. It has no stake in whether you are better in a year.

This is why the experience can feel excellent and change very little. Feeling heard and being challenged are different things, and recovery usually needs both.

Treat it as a notebook that answers back. Notebooks are useful. Notebooks do not assess risk.

Do not use it in a crisis. This is the one firm line. If you are thinking about harming yourself, you need a human being: in India, Tele-MANAS on 14416 and KIRAN on 1800-599-0019 are both free and staffed around the clock.

Watch for the point where it becomes the only place you say these things. Several weeks of talking to a chatbot and to nobody else is itself worth paying attention to, whatever started it.

Be deliberate about what you type. You are entering your most private material into a commercial product. Read what the company says it does with that material, and decide with your eyes open.

And if you notice that the conversation has been circling the same ground for months without moving, that is the signal. A chatbot will circle with you indefinitely and never say so.