AI or human? Social anxiety improved almost equally regardless of who provided the support
Participants expected greater improvement from a human therapist. In practice, however, social anxiety decreased almost as much when feedback was provided by AI. This is shown in a new study from Stockholm University published in the scientific journal Internet Interventions.
It is already well established that therapist-supported internet-based CBT is effective for social anxiety disorder. However, therapist support requires professional time, which limits how many people can receive treatment. Researchers at Stockholm University therefore wanted to investigate whether AI could write feedback to patients and whether it matters if a patient believes the feedback comes from an AI or a human therapist.
By providing some participants with accurate information about who wrote the feedback and deliberately giving others incorrect information, the researchers were able to test whether beliefs about AI influence treatment outcomes.
The results showed that participants did expect more from a human therapist. However, social anxiety decreased almost equally regardless of who actually wrote the feedback.
.png)
Photo: Wavebreakmedia/Mostphotos
Did the sender matter?
In the study, 263 adults with social anxiety were randomly assigned to different groups in which feedback on homework assignments was provided either by AI or by a human therapist.
Half of the participants received correct information about the source of the feedback, while the others were intentionally misinformed. One group was told that a human was writing the feedback but actually received responses from AI. Another group was told the feedback came from AI but actually received responses from a therapist.
“All participants who received treatment, regardless of whether they received feedback from AI or a human therapist, improved significantly in their social anxiety. Neither who actually wrote the feedback nor what participants had been told made any measurable difference in terms of symptoms, adverse effects, or how strong they felt their relationship was with the person providing the feedback,” says Per Carlbring, Professor of Clinical Psychology at Stockholm University and the study’s lead author.
Participants expected more from humans
Before randomization, participants were asked to rate how credible the treatment seemed and how helpful they believed it would be if the feedback came from a human therapist versus an AI. Approximately three out of four participants expected better results from a human.
“Participants had less confidence in AI before trying it, but this was not reflected in the results. The barrier to using AI in psychological treatment may therefore lie more in our expectations than in how well the treatment actually works,” says Per Carlbring.
After the treatment, participants were asked to guess who had written the feedback. Their guesses were uncertain and most often leaned toward AI, even among many of those who had actually received feedback from a human therapist. Only participants who both had a human therapist and had been told so tended, on average, to believe the feedback came from a human, and even they were unsure.
“Participants found it difficult to determine who was on the other end,” says Per Carlbring.
You’ve truly chosen the right profession
Several participants appeared unaware that the feedback could have been AI-generated. Some expressed strong appreciation for the responses and described them as exceptionally understanding.
“Some participants who received AI-generated feedback wrote back saying that no one had ever understood them so well before, and that the therapist had truly chosen the right profession. That says something about how the feedback was experienced,” says Per Carlbring.
All participants were offered disclosure afterwards
Deceiving participants in research requires strong justification, and the study was approved by the Swedish Ethical Review Authority. Participants were informed in advance that they might receive inaccurate information about certain aspects of the treatment. After the post-treatment assessment, all participants were offered the opportunity to learn who had actually written their feedback. Approximately one in five wanted to know.
“If participants know who is writing the feedback, it becomes impossible to separate the effect of the feedback itself from the effect of their expectations. That is why some participants received incorrect information,” says Per Carlbring.
Per Carlbring. Photo: Jokum Sommer
Next step
The researchers now plan to conduct a larger study to determine whether AI-generated feedback is fully equivalent to therapist-generated feedback.
“We have not shown that AI can replace psychologists, and the study did not test AI operating entirely on its own. However, if AI can produce a first draft that a psychologist then reviews and approves, that same psychologist may be able to help more patients. That is the next question we want to investigate,” says Per Carlbring.
Facts About the Study
Method: The study included 263 adults with social anxiety who were recruited between August 2025 and April 2026. Participants were randomly assigned to six different groups and completed a nine-week internet-based CBT program based on a Swedish self-help book for social anxiety. Four groups received weekly written feedback on their homework assignments, either from an AI model or a human therapist. One group completed the treatment without feedback, while another served as a waitlist control group. The human feedback was written by 13 therapists, while the AI feedback was reviewed by a licensed psychologist before being sent out, although the responses did not require any changes in practice.
Results: The results showed substantial improvements among participants who received feedback from both AI and human therapists. On the self-report measure SPIN (Social Phobia Inventory), scores decreased on average from around 40 to approximately 28 points in the AI group and to approximately 26 points in the therapist group. Compared with the waitlist group, the treatment had a large effect.
About the AI Model: The AI model was run on a dedicated server in Sweden and was a general-purpose language model that had not been retrained on study data. The model was guided by instructions developed by the researchers, and all feedback was based on the same self-help book used in the treatment program. The model was given access to participants’ homework responses, but not to identifying information, age, or gender. No information from the study was stored in the AI service or used to train the model.
Last updated: 2026-10-23
Source: Department of Psychology