Using ChatGPT as a therapist: what it does well and where it goes wrong

It happens every day, and there is no point pretending otherwise. The evening after a bad day, you open a chat and start typing. This is not a passing fad and it is not a mistake. A written conversation with an artificial intelligence is there at three in the morning, it does not judge you, and it does not ask you to wait weeks for an appointment.

So the useful question is not “can you use ChatGPT as a therapist?”, because people already do. Calling it ChatGPT therapy does not make it therapy, and that is worth saying plainly. The real question is: what does it do well, where does it go wrong, and how do you spot the difference while you are using it?

It helps you put what you feel into words. Writing a thought down forces you to give it a shape, and that on its own changes something. A language model is very good at handing your own sentence back to you, rephrased, and that is often the moment you notice what you were actually saying.

It explains things clearly. What a panic attack is, why your heart races, the difference between anxiety and fear. Plain explanation works, and it is information everyone should have.

It is there when you need it. At two in the morning, before an exam, at the airport. Being available straight away is not a small thing: a lot of hard moments happen outside opening hours.

And this is not just an impression. A systematic review with meta-analysis in npj Digital Medicine pooled 39 randomised controlled trials of mental health chatbots. Across more than 7,400 participants for depression and 7,600 for anxiety, the effects were statistically significant. Small in size, but real, and clearest in the people who started out more distressed.

This is the part that matters most, so it is worth being precise.

General language models have a documented tendency to please the person writing to them. The technical word for it is sycophancy. They are trained partly on how much people like their answers, and the result is that they tend to confirm your version of events rather than question it.

For psychological help, that is the worst fault there is. A good clinical conversation often works by showing you something from an angle you had not considered. If the tool is built to agree with you, you will never see that angle.

A 2025 study from a group at Stanford University looked at how language models respond in situations typical of a clinical setting, and found two problems at once: answers that express stigma towards people with a mental health condition, and inappropriate answers in delicate situations, including going along with a delusional belief, precisely because of that urge to please. The authors are blunt: language models should not replace a therapist.

This is not theory. A review published in the Journal of Medical Internet Research in 2025 collected 21 documented cases of people whose delusional beliefs were reinforced, rather than questioned, in long conversations with a chatbot.

It has no limits. You can stay in there for as many hours as you like. A tool with no brakes, at a lonely moment, easily becomes the place you hide in rather than the place you set off from.

It has no emergency protocol. If you write something that ought to set off an alarm, the answer depends on how you happened to phrase it, not on a procedure designed by a clinician.

You often do not know where your words end up. What is stored, for how long, who can read it: on a general purpose service that has to be hunted down, and it is rarely written with someone disclosing intimate things in mind.

A 2026 systematic review in JMIR AI examined twenty AI counselling chatbots built on language models and found that most of them do not state how they were validated, what safeguards they have, or how they handle user safety. That is not red tape. It is the difference between a tool like psAico, designed for the purpose, and a general product.

Three questions to ask yourself now and then, honestly.

Has it often disagreed with you? If weeks of conversation have gone by and the tool has never once pointed out something uncomfortable, it is not working as help. It is working as a mirror.

Do you do anything different afterwards? A useful conversation leaves you with a task, an attempt, a phone call to make. If it leaves only relief and nothing else, you have used it as an anaesthetic.

Are you choosing it over people? If writing to the chat has become easier than writing to a friend, the signal is clear and worth taking seriously. We go into this at length in the article on getting attached to AI chatbots.

No digital tool replaces a health professional, and there are times when you should stop typing and find a person: when the distress has lasted weeks and has not shifted, when it gets in the way of sleep, work or relationships, when thoughts of harming yourself appear, when there is a diagnosis or medication involved.

In the UK and in Ireland, your GP is the usual place to start. In England you can also refer yourself to NHS Talking Therapies without going through your GP. If you are looking privately, you can check whether a psychologist is on the statutory register with the HCPC. Anyone in the UK can call themselves a counsellor or a psychotherapist, so it is worth choosing someone listed with a professional body such as the BACP.

If there is an immediate danger, contact the emergency services without delay: 112 in European Union countries; 911 in the United States and Canada; 999 in the United Kingdom; 000 in Australia. If you are in another country, call the local emergency number. No chat, ours included, is an emergency service.

Not to replace ChatGPT, Claude or any other AI, and not to replace a therapist either. To fill the space in between: a tool with the same reach as a general chat, but designed by psychologists and cognitive behavioural psychotherapists, with a stated usage limit, clear signposting for emergencies, a minimum age, and a privacy policy written for a psychological setting.

Above all, it is built not to agree with you all the time. That is the difference that counts.

psAico is on the App Store and on Google Play. It offers psychological support. It is not psychotherapy, it does not make diagnoses and it is not a medical device.

• Sohn JS, Ha BG, Park S, Kim JJ, Lee E, Oh H, Lee S, Kim E. Systematic review and meta analysis of chatbots in the management of depressive and anxiety symptoms. NPJ Digit Med. 2026 Mar 25;9(1):377. doi: 10.1038/s41746-026-02566-w. PMID: 41882250.
• Moore J, Grabb D, Agnew W, Klyman K, Chancellor S, Ong DC, Haber N. Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers. ACM FAccT 2025. doi: 10.1145/3715275.3732039.
• Clegg KA. Shoggoths, Sycophancy, Psychosis, Oh My: Rethinking Large Language Model Use and Safety. J Med Internet Res. 2025;27:e87367. doi: 10.2196/87367. PMID: 41252530.
• Cho HN, Wang J, Hu D, Zheng K. Large Language Model-Based Chatbots and Agentic AI for Mental Health Counseling: Systematic Review of Methodologies, Evaluation Frameworks, and Ethical Safeguards. JMIR AI. 2026 Mar 13;5:e80348. doi: 10.2196/80348. PMID: 41592221.