Research what it means for AI to help

Yes. Research has already found improvements in psychological difficulties with specific AI interventions. But when we want to know what that means for our own conversation, we need to distinguish what people in the study worked with and what their results were compared against. A specialised programme used by a research team with supervision is a different situation from ordinary ChatGPT opened at home.

In the overview that follows, let us therefore keep three questions in view: Who did the tool help? What was it compared with? And under what conditions was it used? These differences explain why the results of individual studies can vary.

What the individual studies examined

Therabot: in a 2025 study with 210 adults, a specialised chatbot achieved better outcomes for the difficulties measured than a waiting-list control group. After responses were sent, the research team checked them and could intervene. It was neither a direct comparison with a psychotherapist nor a test of an ordinary child’s account.^[1]

Kai: a 2026 study with 995 university students in Israel compared a twelve-week AI intervention with group therapy and waiting. For anxiety, AI performed better than both groups; for depression, after correction, the difference was confirmed against waiting, not against group therapy. The platform allowed clinicians to intervene, people in acute crisis were excluded and some authors had ties to the manufacturer.^[2]

Ordinary ChatGPT: a three-week pilot by a Czech team had 147 English-speaking adults at the initial measurement. In an exploratory comparison, depressive symptoms fell more with ChatGPT than in the group without an intervention. This comparison was neither preregistered nor corrected for multiple testing; differences in anxiety and well-being were not confirmed. The finding is preliminary, but does not concern only a special therapeutic product.^[3]

AI application versus CBT materials: in another study with 540 adults, digital records showed participants opening the AI application more often and using it for longer. Overall, however, significantly better outcomes in anxiety and depression were not found. The digital records also did not capture all work with printed materials.^[4]

What broader reviews say

A broader review of 31 randomised studies included 29,637 participants, but the pooled result for psychological difficulties came from 21 studies with 5,929 people. It found a small to moderate average improvement. It brought together different chatbots and populations broadly defined around ages 15–39; generative systems made up only a small part and their overall effect remained uncertain. The result cannot automatically be applied to ordinary ChatGPT or all children.^[5]

A systematic review of 66 studies of general-purpose language models found predominantly simulated and retrospective research; seven studies were prospective. The authors rated the certainty of the evidence mostly as low and found no support for routine unsupervised clinical work, particularly diagnosis, crises and therapeutic interactions.^[6]

What this means for your use

For our own use, research can give us hope of specific help and a clearer idea of what has already been tried. These studies did not, however, test whether ‘a good prompt beats an average therapist’. Your experience with AI may nevertheless be better than a previous experience with a person. It deserves attention in its own right: what helped this time, what you lacked before and what you need next. One person’s experience and a comparison of groups answer different questions.

Ask what the conversation brought you. Relief while talking, a better understanding of the situation, a change the next day or a step you took? All these outcomes can have value, but each means something different. A pleasant evening need not be treatment to be worthwhile. But if long-term difficulties are not improving, repeated relief alone is not a reason to postpone further help.