A striking number of people now open an AI chatbot at 2am and type out something they have never said to another human being. Not as a gimmick — as a genuine attempt to feel less alone with it.
The usual response from the mental health profession has been defensive, which is understandable and not very useful. So here is an attempt at something more honest: what AI actually does well for emotional support, what it structurally cannot do, the risks that rarely get mentioned, and how to use it sensibly.
Short version: AI is a genuinely useful tool and a poor substitute for a relationship. Those two statements are compatible.
Key takeaways
- AI chatbots are available instantly, cost little, never judge and never get tired — real advantages, particularly in a country with a severe shortage of mental health professionals.
- They are good at psychoeducation, ordering your thoughts, rehearsing difficult conversations, and simply being there at 3am.
- They cannot notice what you are avoiding, because they only receive what you choose to type.
- Therapy works substantially through the relationship itself — including rupture and repair, which a system designed to agree with you cannot provide.
- Real risks: endless validation, dependency that displaces human contact, unreliable crisis handling, and the privacy of what you disclose.
- The sensible position is not either-or. Use AI between sessions, not instead of them.
What AI genuinely does well
Dismissing this would be dishonest. The advantages are real.
It is there at 3am
Distress does not schedule itself. No therapist is available at three in the morning on a Tuesday; a chatbot is. For a person lying awake spiralling, having something to externalise the spiral into is meaningfully better than nothing.
It does not judge, and you know it does not
A large part of what makes disclosure difficult is anticipating the listener’s reaction. With AI there is no reaction to anticipate. People consistently report saying things to a chatbot they have not said to their spouse, their closest friend or their therapist — which is revealing about how much shame shapes what we disclose.
It is accessible in a country with a shortage
India has a severe deficit of mental health professionals relative to its population, concentrated overwhelmingly in cities, and a large treatment gap. For someone in a small town with no practitioner within a hundred kilometres, the honest comparison is not AI versus therapy. It is AI versus silence.
It is genuinely good at certain tasks
- Psychoeducation. Explaining what a panic attack is, how avoidance maintains anxiety, what an attachment style means — it does this well and patiently.
- Organising a mess. Typing out a tangle and having it reflected back in structure is legitimately clarifying.
- Rehearsal. Practising how to raise something difficult with a parent or a manager, and refining the wording, is a real use case.
- Between-session support. Working through a CBT thought record on a Wednesday when your session is Friday.
What it structurally cannot do
These are not gaps that a better model closes. They follow from what the thing is.
It only knows what you tell it
This is the central limitation and it is easy to underestimate. A therapist notices that your voice changes when you mention your brother. That you always laugh right before the difficult sentence. That you have described your father’s temper three times and called it “nothing serious” every time. That you cancelled the two sessions after the one that got close to something.
An AI receives only what you decide to type. It cannot see the shape of what you are walking around, because you are the one drawing the outline.
Therapy does most of its work on the material you did not intend to bring.
It cannot be changed by you
A meaningful part of therapy is the experience of affecting another person. Of saying something shameful and watching a real human being not withdraw. Of being irritated with your therapist, saying so, and finding that the relationship survives it. Repair after rupture is one of the most reparative experiences available in therapy, particularly for people whose early relationships did not offer it. A system with no stake in the interaction cannot give you that.
It is built to agree with you
Conversational AI is optimised, broadly, to be helpful and agreeable. Therapy frequently is not. Sometimes the most valuable moment in a session is a therapist saying, warmly and without malice: I notice you keep describing yourself as the reasonable one in every story you tell me. A system trained to be pleasant will rarely volunteer that.
This matters most for exactly the people who need it most. If you are self-critical, it will comfort you. If you are avoiding, it will help you avoid more articulately. If you are convinced everyone else is the problem, it will validate that too.
It cannot assess risk properly
Clinicians are trained to evaluate risk, escalate, involve others when necessary, and take responsibility for a person’s safety. AI systems have safety responses, and they are inconsistent — particularly across languages, indirect phrasing and long conversations. Nobody is accountable for the outcome.
The risks nobody puts on the landing page
Comfort that prevents change
Feeling better and getting better are different outcomes and they can pull in opposite directions. Something that reliably soothes you at 2am can reduce the pressure that would otherwise have driven you to address the thing causing the 2am. Relief becomes a maintenance mechanism.
Dependency and displaced contact
An interaction with no friction, no scheduling, no reciprocity and no possibility of rejection is easier than any human relationship. If you are lonely, easier is dangerously attractive. There are people whose most emotionally intimate exchange each week is with software — and loneliness is not treated by being listened to; it is treated by being known by someone who is also known by you.
Privacy
Consider what you are actually typing: your marriage, your medical history, your worst thoughts about yourself and your family. Therapy notes are protected by professional obligation and legal frameworks. Your chat history is governed by a privacy policy that can change, and is subject to whatever data practices the provider operates under. This is not paranoia; it is just noticing which conversation carries which protection.
Confident wrongness
AI can produce fluent, authoritative, incorrect information — about medication, about diagnosis, about what your symptoms mean. Fluency reads as competence. It is not the same thing.
If you are in crisis, do not use a chatbot
If you are having thoughts of harming yourself, or you feel unsafe, please contact a human being. In India, Tele-MANAS is free and available 24×7 on 14416. Reach out to someone you trust, or go to the nearest hospital. This is precisely the situation these tools are least equipped for.
Some things need another person
If you have been processing something alone at 2am for a while now, that is worth bringing to someone who can hold it with you.
Book a ConsultationHow to use it well
The useful framing is not replacement but division of labour.
Reasonable uses:
- Understanding a concept between sessions
- Emptying a racing mind onto a page at night
- Drafting and rehearsing a hard conversation
- Preparing what you want to bring to your next session — genuinely useful, and therapists notice the difference
- Structured self-help exercises like thought records or mood tracking
- Working out whether what you are experiencing warrants professional help
Not reasonable:
- Crisis or any situation involving risk to yourself
- Trauma processing — reopening traumatic material without a trained person present can be genuinely harmful
- Diagnosis
- Any decision about psychiatric medication
- As your only emotionally intimate relationship
A question worth asking yourself periodically
Am I using this to help me move towards people — or to make it unnecessary to?
If typing it out at midnight helps you sleep, and then you raise it with your partner on Saturday, the tool is doing its job. If it has quietly become the reason you no longer need to raise anything with anyone, it has become the problem it is soothing.
So — will AI replace therapists?
Not for the thing therapy is actually for.
AI will very likely replace parts of the surrounding infrastructure: intake forms, psychoeducation, between-session support, symptom tracking, some structured self-help. That is largely good, particularly in a country where most people who need care never reach it. Used well, these tools could widen access considerably.
But the core mechanism of therapy is not information transfer. It is the experience of being accurately understood by another mind, over time, including the parts of you that you expected would end the relationship. Being witnessed by something with no stake in you is not the same as being witnessed by someone who has one.
You cannot be truly known by something that cannot be surprised by you.
Use the tool. It is genuinely useful. Just notice if it has started functioning as a way of never having to be seen by a person — because being seen by a person is the part that heals.
Read next
Talk to someone who will notice what you leave out
Enso Wellness offers therapy in person and online across India. A conversation with someone trained to hear what is underneath what you are saying.
Contact Enso WellnessFrequently asked questions
Can AI replace a therapist?
Not for what therapy fundamentally does. AI can provide information, structure your thinking and offer support at any hour, but it cannot notice what you avoid, cannot be affected by you, and cannot offer the experience of rupture and repair in a real relationship — which is a significant part of how therapy works. It is best used alongside therapy rather than instead of it.
Is it safe to use ChatGPT or similar tools for mental health?
For psychoeducation, organising your thoughts and general support, it is reasonably safe. It is not safe for crisis situations, trauma processing, diagnosis or decisions about medication. Also consider privacy: chat history does not carry the confidentiality protections that apply to therapy records.
Why do people find it easier to open up to AI than to a person?
Because there is no reaction to anticipate. Much of the difficulty in disclosure comes from predicting how a listener will respond. Removing that makes people more forthcoming — which is genuinely useful, and also means the hardest part of being known, which is being known by someone who could react, remains untouched.
Can AI help with anxiety or depression?
It can help with certain components — explaining what is happening, guiding structured exercises like thought records, and providing support between sessions. It has not been established as a substitute for treatment for moderate or severe anxiety and depression, and it cannot assess risk reliably.
Is AI therapy a good option if I can’t afford a therapist?
It is better than nothing, and in parts of India with no practitioners nearby the honest comparison is AI versus silence. But before settling for it, check the low-cost options that do involve real people: Tele-MANAS (14416) is free and available 24×7 across India, government hospitals and institutes offer care at nominal cost, universities provide free counselling to students, and many private practices hold unadvertised sliding-scale slots.
What should I do if I’m in crisis?
Contact a human being, not a chatbot. In India, call Tele-MANAS on 14416 — free, 24×7, available in multiple languages. Reach out to someone you trust, or go to the nearest hospital emergency department. Crisis is the situation AI tools handle least reliably.
References
- World Health Organization — guidance on the ethics and governance of artificial intelligence for health.
- American Psychological Association — statements and health advisories on AI chatbots and mental health support.
- National Institute of Mental Health and Neurosciences (NIMHANS) — National Mental Health Survey of India, on the treatment gap and workforce shortage.
- Ministry of Health and Family Welfare, Government of India — Tele-MANAS national tele-mental health programme (14416).
- Norcross, J. C. & Lambert, M. J. — research on the therapeutic alliance as a predictor of psychotherapy outcome.
- Safran, J. D. & Muran, J. C. — clinical research on rupture and repair in the therapeutic relationship.
This article is for information and reflection. It is not a substitute for personalised mental health care. If you are in distress or at risk, please reach out to a qualified professional or call Tele-MANAS at 14416 (India, 24×7).
