AI feedback for social anxiety was linked to symptom improvements similar to those seen with therapist-written messages in a Swedish clinical trial announced on 20 October. But the researchers could not establish that AI was as effective as a human therapist, and a psychologist reviewed every AI-generated reply before it reached a participant.
Researchers at Stockholm University randomly assigned 263 adults with social anxiety to six groups in a nine-week study of internet-based cognitive behavioural therapy (CBT). The research, published in the journal Internet Interventions, examined both who wrote patients’ weekly feedback and what patients believed about its author.
How AI feedback for social anxiety was tested
Four groups followed an online treatment programme and received written comments on their exercises roughly once a week. In two groups, the comments were written by a human therapist; in the other two, they were generated by AI. Some participants were accurately told who wrote their feedback, while others were deliberately given the opposite information. A fifth group completed the programme without written feedback, and a sixth was placed on a waiting list.
The design allowed the researchers to separate the effects of AI-written versus therapist-written feedback from expectations about its source. According to study co-author Per Carlbring’s account of the trial, participants were warned beforehand that they might receive misleading information about some aspects of the treatment, and the study had ethics approval.
The AI was a general-purpose language model, not a chatbot conducting therapy independently. It ran on a server in Sweden and followed detailed instructions based on previous research and the self-help book used in the programme. It received patients’ homework responses but not identifying details, age or gender. The researchers said the texts were not retained by the AI service or used to train the model. A licensed psychologist checked the AI replies before they were sent and, according to the university, rarely needed to change them.
Similar improvements did not prove equal effectiveness
Participants receiving treatment generally reported substantial reductions in social anxiety. The researchers found no measurable difference attributable to the feedback’s actual author or to what participants were told about that author. They also reported no detectable differences in negative effects or patients’ perceived relationship with the person providing feedback, as Swedish research news service forskning.se reported.
Expectations had nevertheless favoured human therapists. Before assignment, 74% of participants rated human-written feedback as the more credible option, compared with 14% who favoured AI, according to Carlbring.
That finding does not establish that the methods are interchangeable. The improvement was numerically greater in the therapist-feedback groups, and the trial did not meet its predefined threshold for showing that AI was not meaningfully worse. The uncertainty around the estimates still allowed for a clinically relevant difference. In other words, the study did not detect a clear difference, but it could not prove equivalence or non-inferiority.
The role of human supervision
Per Carlbring, professor of clinical psychology at Stockholm University, emphasised the limits of the findings in a university statement:
“We have not shown that AI can replace psychologists, and the study did not test AI working entirely on its own.”
The results apply to this structured online treatment, this group of adults and the period studied. They do not establish the long-term effects of AI-written feedback or the safety of unsupervised mental health chatbots.
The practical question now is whether AI can prepare useful first drafts that professionals approve, reducing the time needed for routine feedback without compromising care. That potential saving, rather than the replacement of psychologists, remains to be tested.
