Research Methods in Psychology -- Impact of Therapist Type and Gender on Perceived Empathy in Psychotherapy

University: James Cook University

Word Count: approximately 2,100 words

Abstract

This study investigated whether perceptions of therapist empathy are influenced by knowledge of whether a therapist is human or AI-based, and whether therapist gender affects these perceptions. Participants (N = 313) read identical therapy transcripts but were randomly assigned to conditions describing the therapist as human, an AI chatbot, or a human-supervised AI chatbot, with therapist gender also manipulated. A 3 (therapist type) x 2 (therapist gender) between-subjects factorial ANOVA revealed a significant main effect of therapist type on perceived empathy, F(2, 307) = 21.03, p < .001, with human therapists and human-supervised AI therapists rated significantly more empathetic than unsupervised AI therapists. No significant effect was found for therapist gender, and no significant interaction emerged. These findings suggest people show an anthropocentric bias against AI-delivered psychotherapy resembling the bias previously observed in perceptions of AI-created art, potentially creating barriers to AI acceptance in therapeutic contexts.

Introduction

AI-based chatbots are increasingly used to deliver psychological treatments, potentially expanding access to therapy by reducing wait times and costs, but a central concern is whether AI therapists can provide the empathetic connection fundamental to effective therapeutic relationships. Prior research demonstrating a systematic bias against AI-created art -- where identical artworks were rated less emotionally moving and creative when labelled AI-generated (Millet et al., 2023) -- suggests a broader anthropocentric bias that might similarly affect perceptions of AI therapists. Therapist gender has separately been shown to influence therapeutic perceptions in traditional settings (Bhati, 2014), raising the question of whether gender effects persist, interact with, or are overshadowed by the human-AI distinction. The study also draws on evidence that public acceptance of AI increases when humans remain "in the decision loop" (Aoki, 2021), motivating inclusion of a human-supervised AI condition. Three hypotheses were tested: human therapists would be perceived as more empathetic than AI therapists (H1); human-supervised AI therapists would be perceived as more empathetic than unsupervised AI (H2); and female therapists would be perceived as more empathetic than male therapists regardless of human/AI status (H3).

Method

Design. A 3 (therapist type: human vs. chatbot vs. human-supervised chatbot) x 2 (therapist gender: male vs. female) between-subjects factorial design was used, with perceived empathy as the dependent variable. An a priori power analysis (G*Power) based on Millet et al.'s (2023) medium-to-large effect sizes indicated a required sample of 158 for a medium effect (f = 0.25, alpha = .05, power = .80); a target of approximately 300 was set to account for data loss and detect smaller effects.

Participants. The final sample comprised 313 participants recruited from the Australian and Singaporean campuses of James Cook University via campus bulletin boards, the SONA psychology participant pool, and staff emails, including students, staff, and other community members.

Materials. All participants read an identical ~2,000-word therapy transcript between a therapist (Tom or Tara, depending on gender condition) and a fictional male client, Howard, seeking help for problem gambling; the transcript was initially generated using AI and then edited by the research team, with only the framing information about the therapist varied across conditions. Therapist type was manipulated via introductory text describing the therapist as human, as an unsupervised AI ("HelpBot"), or as an AI operating "under the supervision of a human therapist, who periodically reviews transcripts." Therapist gender was manipulated by naming the therapist Tom or Tara, with AI conditions describing the AI as "adopting a persona" of that gender.

Measures. Perceived empathy was measured with a modified three-item scale adapted from Hu et al. (2022), summed into a total perceived empathy score.

Procedure. Following JCU Human Research Ethics Committee approval, data was collected online via Qualtrics with random assignment to one of six conditions. Participants read the therapist description and transcript, then completed the empathy measure and demographic questions; the study took approximately 15-20 minutes, with student participants receiving course credit.

Results

Human therapists received the highest empathy ratings (M = 12.44, SD = 2.45), followed by human-supervised chatbot therapists (M = 11.96, SD = 2.69), with unsupervised chatbot therapists rated lowest (M = 10.11, SD = 2.98). Female (M = 11.58, SD = 2.83) and male (M = 11.49, SD = 2.94) therapists received similar ratings.

Shapiro-Wilk tests indicated normality violations across all three therapist-type groups, though ANOVA is considered robust to this with large samples; Levene's test confirmed homogeneity of variance was maintained, F(5, 307) = 1.53, p = .180.

The factorial ANOVA revealed a significant main effect of therapist type, F(2, 307) = 21.03, p < .001, partial eta-squared = .12, but no significant main effect of therapist gender, F(1, 307) = 0.08, p = .781, partial eta-squared < .001, and no significant interaction, F(2, 307) = 1.75, p = .176, partial eta-squared = .01.

Tukey's HSD post hoc tests showed human therapists were rated significantly more empathetic than unsupervised chatbot therapists (mean difference = 2.33, p < .001, 95% CI [1.44, 3.21]), and human-supervised chatbot therapists were rated significantly more empathetic than unsupervised chatbot therapists (mean difference = 1.85, p < .001, 95% CI [0.96, 2.74]). The difference between human therapists and human-supervised chatbot therapists was not significant (mean difference = 0.47, p = .410, 95% CI [-0.40, 1.35]).

Discussion

Principal findings and theoretical implications. As hypothesised (H1), human therapists were perceived as significantly more empathetic than AI therapists despite identical transcript content, aligning with Millet et al.'s (2023) finding of bias against AI-created art and suggesting a broader anthropocentric bias extending to empathy perception in therapeutic contexts. H2 was also supported: human-supervised AI therapists were perceived as significantly more empathetic than unsupervised AI, consistent with Aoki's (2021) finding that keeping humans "in the decision loop" increases AI acceptance -- notably, human supervision closed the gap with human therapists entirely, with no significant difference between the two conditions. Contrary to H3, therapist gender had no significant effect and did not interact with therapist type, which the study suggests may reflect the human-AI distinction overshadowing gender effects, or reduced gender salience in written-transcript format compared to face-to-face interaction.

Limitations and future directions. Participants read a static transcript rather than experiencing an actual session, which may not capture the nuances of live therapeutic interaction; future research could use video or live interaction stimuli. Prior attitudes toward AI were not measured as a potential moderator. The normality violation, while likely tolerable given sample size, suggests non-parametric analyses could be used in future replications. The university-based sample may limit generalisability to populations who could benefit most from AI-expanded mental health access.

Practical implications. The findings suggest technically competent AI alone may be insufficient for user acceptance; developers must also address perceptual barriers. Since human supervision substantially improved perceptions of AI therapists, positioning AI as an augmentation to human therapy rather than a standalone replacement may be a practical path to improving acceptance and perceived effectiveness.

Conclusion

Perceptions of therapist empathy are significantly shaped by knowledge of whether a therapist is human or AI, even when the therapeutic content itself is identical -- a bias mirroring that observed in perceptions of AI-created art. Human supervision appears to substantially mitigate this bias, offering a potential pathway for the effective integration of AI into mental health services as the field continues to expand.

References

Aoki, N. (2021). The importance of the assurance that "humans are still in the decision loop" for public trust in artificial intelligence: Evidence from an online experiment. Computers in Human Behavior, 114, Article 106572. Aoki, 2021 Atkan, M. E., Turhan, Z., & Dolu, I. (2022). Attitudes and perspectives towards the preferences for artificial intelligence in psychotherapy. Computers in Human Behavior, 133, Article 107273. Atkan et al., 2022 Barnett, A., Savic, M., Pienaar, K., Carter, A., Warren, N., Sandral, E., & Lubman, D. I. (2021). Enacting 'more-than-human' care: Clients' and counsellors' views on the multiple affordances of chatbots in alcohol and other drug counselling. International Journal of Drug Policy, 94, Article 102910. Barnett et al., 2021 Bhati, K. S. (2014). Effect of client-therapist gender match on the therapeutic relationship: An exploratory analysis. Psychological Reports, 115(2), 565-583. Bhati, 2014 He, L., Basar, E., Wiers, R. W., Antheunis, M. L., & Krahmer, E. (2022). Can chatbots help to motivate smoking cessation? A study on the effectiveness of motivational interviewing on engagement and therapeutic alliance. BMC Public Health, 22(1), Article 726. He et al., 2022 Hu, T., Xu, A., Liu, Z., You, Q., Guo, Y., Sinha, V., Luo, J., & Akkiraju, R. (2022). Touch your heart: A tone-aware chatbot for customer care on social media. Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, 1-13. Hu et al., 2022 Millet, K., Buehler, F., Du, G., & Kokkoris, M. D. (2023). Defending humankind: Anthropocentric bias in the appreciation of AI art. Computers in Human Behavior, 143, Article 107707. Millet et al., 2023