All participants first completed baseline questionnaires and then heard a recording lasting 2 minutes and 8 seconds stating that psychotherapy does not help most people over the long term. They were also asked to imagine a time when they had felt depressed and had accomplished less than usual. This standardized induction reduced expectations of treatment before the experimental videos began, with a separate pretest confirming the change. Participants completed the first expectation assessment after the audio and were then told to imagine attending an online therapy session for depressive symptoms. The researchers excluded the earlier baseline measurement from the primary analysis so that it would focus on whether the subsequent videos could reverse the induced negativity.

The 484 analyzed participants had a mean age of 34.43 years, with a standard deviation of 12.02 and a range of 18 to 74 years. The sample included 259 women, 221 men and four gender-diverse people. Recruitment occurred through university mailing lists and the Prolific online platform between March and August 2025. Eligibility criteria required adequate language skills, no hearing or visual impairment and no current psychotherapy; people reporting psychotic disorders or severe neurological disorders were excluded. The sample was relatively well educated and had a low burden of depressive symptoms: mean Patient Health Questionnaire-9 scores, on a scale ranging from 0 to 27, varied from 6.20 to 7.59 across the four groups. These characteristics limit how directly the findings can be applied to people receiving treatment for depression.

Participants were randomly allocated to four sequences: human therapist followed by human therapist, human followed by avatar, avatar followed by human, or avatar followed by avatar. The first video, lasting 1 minute and 8 seconds, presented a female therapist acknowledging that the imagined participant’s recent weeks might have been difficult and lonely. The second, lasting 1 minute and 19 seconds, conveyed the message that new studies show current depression treatments are quite effective and help most people. The human and avatar versions used the same person, appearance, script and voice. Participants were told before each video whether the therapist would be human or an avatar. This design kept visual and verbal material constant while testing the effect of participants’ beliefs about the therapist’s identity, although the avatar’s nonverbal performance introduced a separate limitation discussed below.

The avatar was generated with HeyGen software and designed as a photorealistic replica of the real therapist. At first glance, it could appear human, but closer inspection revealed repetitive facial expressions and hand movements. Its expressions were not synchronized with the spoken words and instead appeared randomly. Pretests involving two candidate therapist-avatar pairs, with 22 participants in each group, found that one actor received higher empathy ratings overall; that actor and avatar were therefore selected for the main experiment. This design strengthened the comparison between the disclosed identity of the therapist and the stimulus itself. It did not, however, allow the researchers to cleanly distinguish the effect of believing a therapist was an avatar from the effect of the specific avatar’s weaker and less synchronized facial and gestural behavior.

Outcome expectations were assessed with the German Credibility Expectancy Questionnaire, which contains four nine-point Likert items and two percentage items transformed to a nine-point scale for analysis. Its internal consistency was good, with a Cronbach’s alpha of 0.91. Perceived therapeutic alliance was measured after each video with the 12-item German Working Alliance Inventory Short Version, ranging from 12 to 60 points, with an alpha of 0.93. The first video also produced higher perceived empathy when delivered by the human therapist than by the avatar, a large and statistically significant difference. Warmth ratings, in contrast, did not differ significantly between human and avatar videos. Participants generally rated alliance with the avatar above 40 points on the 60-point scale, indicating that the avatar was not uniformly viewed as an ineffective therapeutic partner, even though the human sequence ultimately received stronger ratings.

Across the experiment as a whole, outcome expectations rose significantly after the two videos compared with their level immediately after the negative-expectation induction. A statistically significant time-by-group interaction showed that the four groups followed different trajectories. The two-human sequence had the strongest final expectations and alliance ratings. Expectations and alliance rose after the second video in the avatar-human and avatar-avatar groups, but the reported final expectation comparisons did not establish that every increase in those groups was smaller than the increase in the human-human group. The human-then-avatar group departed from the expected pattern: its expectations and alliance declined after the avatar delivered the expectation-challenging message. At the final measurement, its increase in outcome expectations from the post-audio assessment was significantly smaller than in the human-human group, as was the increase in the avatar-avatar group.

The working-alliance results further distinguished the sequences. Alliance remained stable between the two videos in the human-human group, whereas the human-avatar group experienced a larger decline, leaving it 4.41 points below the human-human reference group at the final measurement. By contrast, the avatar-human group improved by 1.97 points relative to the human-human sequence. The researchers proposed two overlapping explanations. Avatar-delivered information may have been generally less effective at challenging treatment expectations, while the particular move from a human to an avatar may also have produced a negative contrast effect. The latter possibility is consistent with expectations and alliance falling in the human-avatar group but continuing to improve in the avatar-avatar group. Because nonverbal synchronization and movement were not manipulated independently, the experiment cannot determine whether the pattern arose primarily from participants’ knowledge that the therapist was an avatar or from the avatar’s visibly artificial performance.

A preregistered mediation analysis indicated that alliance after the first video partially accounted for the statistical association between expectations after the negative induction and expectations after the second video. The estimated indirect effect was 0.09, with a 95% confidence interval from 0.05 to 0.14, while a direct association of 0.28 also remained. However, allowing the indirect effect to differ among the four groups did not significantly improve model fit. The researchers therefore cautioned that this cross-group relationship should be understood as a chronological association rather than an explanation for why the four video sequences differed. Therapy motivation also showed a time-by-group interaction, with the human-human sequence producing the largest increase. At the final measurement, the other three sequences showed smaller increases from the post-audio assessment. Neither the experiment’s measured ratings nor its motivation results establish that participants sought or benefited from actual psychotherapy.

Several limitations constrain interpretation. The full procedure lasted about 25 minutes and involved prerecorded videos rather than live interaction, an enduring therapeutic relationship or actual treatment. Of 66 participants who answered an optional open comment field, 16 expressed uncertainty about whether a therapist was human or an avatar. The avatar’s near-realistic appearance may therefore have produced confusion or an uncanny-valley response, although a sensitivity analysis excluding those participants left the direction of the effects unchanged. The empathy video always came first and the expectation-violating video second, so video position was confounded with content. The use of one human-avatar pair, brief exposure, a relatively well-educated and healthy sample, and the avatar’s unsynchronized movements also prevents broad conclusions about clinical practice. Future research would need clinical samples, additional therapist-avatar pairs, systematically varied nonverbal synchrony and separate tests of video order to establish whether the sequence pattern persists under more realistic conditions.

Subject of Research: Experimental effects of human and photorealistic avatar therapists on treatment expectations and therapeutic alliance

Article Title: Photorealistic avatars can help to improve treatment outcome expectations

Article References: Bauermann, P., & Rief, W. (2026). Photorealistic avatars can help to improve treatment outcome expectations. Frontiers in Psychology, 17, Article 1960478. https://doi.org/10.3389/fpsyg.2026.1960478

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1960478

Keywords: photorealistic avatars, psychotherapy, therapeutic alliance, outcome expectations, sequence effects, blended care, digital mental health, empathy, randomized experiment, depression, Photorealistic, avatars