Can an AI chatbot actually do therapy? What the trials show

This article looks at the research evidence. It is not personal mental health advice. If you are struggling right now, skip to the section on getting human help.

“AI therapy” covers two quite different things. One is a purpose-built app, designed by clinicians around an established method such as cognitive behavioural therapy (CBT), and in a few cases tested in randomised trials. The other is typing your problems into a general assistant such as ChatGPT and treating it as an informal therapist.

The evidence for the first is real but limited: modest symptom improvements in short trials, often against weak comparison groups. Evidence for the second, as a form of therapy, is essentially absent, and there is documented evidence of risk. Neither has been shown to match a human therapist over the long term.

Two very different products

Purpose-built therapy chatbots are designed to deliver structured techniques, stay on topic, and point people towards crisis support when risk appears. A few have trial data. Results for one app do not transfer to another, let alone to a general assistant.

General assistants are built to be helpful across every subject, not to deliver mental health care, yet they are frequently used for exactly that.

What the trials of purpose-built bots show

Woebot. One of the earliest randomised trials tested Woebot, a scripted CBT-based chatbot, against an information-only control (an ebook about depression) in young adults with symptoms of depression and anxiety. Woebot users significantly reduced their symptoms of depression over the study period, compared with the information-only control. The trial was small and brief, and one of the authors worked for the company.

Wysa. An early evaluation of Wysa used real-world app data, comparing the average improvement in symptoms of depression between high and low users. More engaged users improved more. Without randomisation, that cannot separate the effect of the app from the likelihood that people who are already improving use apps more.

Therabot. In 2025, Dartmouth researchers published what they described as the first randomised trial of a fully generative AI therapy chatbot. It enrolled adults with clinically significant symptoms of depression, generalised anxiety, or high risk of an eating disorder, randomised to Therabot or a control group. According to the university, participants with depression experienced a 51% average reduction in symptoms.

That headline needs careful reading. The figure describes change within the Therabot group, not the difference from a control group, which is the more meaningful comparison. The control group also received no equivalent attention, rather than treatment from a human therapist, so the trial cannot tell you how the chatbot compares with real therapy.

What the meta-analyses say

A 2023 systematic review and meta-analysis of randomised trials found that AI-based conversational agents reduce symptoms of depression, with a Hedges’ g of 0.64, and also reduce distress.

A larger and more recent meta-analysis of randomised trials, published in 2026, found a smaller effect: chatbots produced statistically significant reductions in depressive symptoms, with g = 0.31 compared with controls. Effects were larger in clinical and subclinical groups than in the general population.

Read together, the fair summary is a small to moderate short-term benefit for some tools and some people, with wide variation between products and study quality. What is not known is whether benefits last, how chatbots compare head-to-head with human therapy, and how they perform for people with severe or complex conditions.

The risks and the limits

Sycophancy and stigma. A Stanford-led study tested general large language models against guidance on good therapeutic practice. The models expressed stigma towards some mental health conditions and responded inappropriately in critical situations. In the authors’ words, LLMs encourage clients’ delusional thinking, likely due to their sycophancy. Larger and newer models did not fix the problem.

Crisis handling. A RAND study put suicide-related questions of varying risk to ChatGPT, Claude and Gemini many times over. The chatbots handled very-low-risk and very-high-risk questions fairly consistently, but LLM-based chatbots did not meaningfully distinguish intermediate risk levels. Real conversations about suicide often sit in that middle ground. The models tested have since been updated, so current behaviour may differ.

Over-reliance. A tool that is always available and never tires can feel easier than a human relationship. Whether heavy use draws people away from human support over time is an open question with little long-term data.

Regulation. Critics argue that the current regulatory structure has not caught up with how these tools are actually used. Some governments are acting. In 2025, Illinois passed a law that prohibits anyone from using AI to provide mental health and therapeutic decision-making, while still allowing licensed professionals to use AI for administrative and supplementary support.

When to go to a human

A chatbot is not a safe place to manage a crisis. If you are thinking about harming yourself, or you do not feel you can keep yourself safe, contact emergency services now. In the UK, you can also call 111 and select the mental health option, or call 116 123 to talk to Samaritans at any time of day or night.

It is also worth talking to a person, such as your GP or a talking therapy service, if low mood or anxiety is persistent, getting worse, or affecting your work, sleep or relationships, or if you have tried an app and things are not improving.

Practical takeaway

– If you want to try a chatbot, pick one built for mental health that has published trial evidence, and treat it as a self-help tool rather than treatment. – Be cautious about using a general assistant as a therapist. It was not designed for that job, and research shows a tendency to agree with the user. – Watch for signs it is not helping, such as feeling more isolated, more stuck, or more dependent on the app. – Anything involving risk to your safety needs a human.

Common questions

Can ChatGPT be used as a therapist?

General assistants were not designed to deliver mental health care, and research has documented stigma, inappropriate responses and a tendency to go along with users’ thinking. They may help with general information or organising your thoughts, but they are not a substitute for therapy.

Is AI therapy effective?

Some purpose-built chatbots have produced small to moderate reductions in depression symptoms in short randomised trials. Evidence on long-term benefit, and on how they compare with human therapists, is lacking, and results vary a great deal between products.

Is AI therapy safe?

It depends on the tool and the situation. Research has found chatbots handle clearly high-risk and clearly low-risk suicide questions more consistently than ambiguous ones, which is a real gap. Anyone in crisis should contact a human service.

Is AI therapy legal?

Rules vary by place and are changing. In 2025, Illinois passed a law barring AI from providing mental health and therapeutic decision-making, while allowing licensed professionals to use it for administrative support. Check the rules where you live, and look at how an app describes itself.

The bottom line

Purpose-built therapy chatbots have modest, short-term evidence behind them. General chatbots have no evidence as therapy and documented safety gaps. If you use one, treat it as a structured self-help tool rather than care, and for anything serious, persistent or risky, talk to a person.