The study, conducted in December 2023 as a preregistered randomized online experiment, investigated how humanized and nonhumanized AI-generated avatars impact intention to comply with postcare instructions in a health care setting (Penn Wharton Credibility Lab 154774). It specifically assessed whether AI-generated messages are as effective as those from a human and explored how disclosure of the message’s AI origin influences this effect.

Participants were told that they had to undergo dental implant surgery during a visit to the dentist, a procedure for which postoperative care is crucial. To encourage them to follow the instructions, they were told that the dentist features an instructional video on post–dental implant surgery care. All materials are available in Multimedia Appendix 1.

Ethics approval was obtained from the research ethics committee of the Faculty of Physical and Mathematical Sciences at the University of Chile (#019). All participants provided informed consent prior to participation by electronically approving a consent form. Data were collected and analyzed anonymously. Participants received compensation for their participation in the study (£0.75-£0.90; £1=US $1.27 as of December 2023). The person featured in the real-person video condition participated voluntarily and provided written informed consent authorizing the use of her image and likeness for research purposes. The experiments were preregistered at the Penn Wharton Credibility Lab (154774 and 164881). Although the Penn Wharton Credibility Lab is not a World Health Organization–accredited trial registry, the requirement for registration in a World Health Organization–accredited registry was waived because the study did not involve a direct medical intervention.

The sample consisted of 650 adults (mean age 44.5, SD 14.82 years; n=329, 50.6% women). Participants were recruited through Prolific Academic (Prolific Academic Ltd), and a preintervention screening ensured that all resided in the United States and English was their first language. Table S1 in Multimedia Appendix 2 present summary statistics for the sample.

Participants were randomly assigned to 1 of 5 video conditions, all delivering identical dental posttreatment instructions. Videos and audio were generated using Synthesia (Synthesia Ltd), an AI-based platform that produces human-like digital avatars delivering scripted health information. The conditions varied by the type of spokesperson—a real person, a humanized AI avatar, or an animated AI avatar—and whether participants were informed that the video was AI generated. All videos featured a White speaker who was a woman. The five groups were as follows: (1) real person (baseline), (2) humanized AI, (3) animated AI, (4) disclosed humanized AI, and (5) disclosed animated AI. The humanized AI avatar was designed to closely approximate a real person’s appearance with near-photorealistic facial features and naturalistic expressions. The animated AI avatar used stylized, nonphotorealistic visual features. All videos delivered identical scripted content. Visual examples are provided in Multimedia Appendix 3 and Multimedia Appendix 4. Participants were not required to watch the entire video, but time spent on the video page was tracked and did not differ across conditions (>97% spent sufficient time to watch the complete video).

The study’s primary outcomes were intention to comply and understanding. Intention to comply was assessed using 2 Likert-scale questions regarding the likelihood of following instructions and returning to the same dentist (ranging from 1 =“extremely unlikely” to 7 =“extremely likely”). Understanding was measured with a 5-item multiple-choice quiz based on the video content, developed with a dentist’s guidance and expertise. The study also tracked whether participants rewatched the video before answering, as they could normally do. Participants also answered additional questions about the video, including its clarity, usefulness, engagement, and the speaker’s voice and appearance. Each of the 4 AI avatar conditions was compared with the “real person” condition, which served as the baseline. Finally, participants provided their opinions on topics related to AI and health care and answered demographic and open-ended questions about comments on the videos and the survey.

As preregistered, we conducted linear regressions with robust SEs for the main outcome variables. Analyses were performed using Stata (version 18; StataCorp).

The first experiment examined whether a humanized AI video has the potential to deliver postcare instructions in health care. In the second randomized experiment, we examined the effect of patient-speaker gender congruence on intention to comply after the same dental procedure described in the first experiment. The study was preregistered (Penn Wharton Credibility Lab 164881), and ethics approval was obtained from the same institution as that for the first experiment.

All 256 participants (mean age 41.3, SD 13.71 years; n=128, 50% women) were residents of the United States and native English speakers recruited via the Prolific Academic panel. A total of 4 participants were excluded from an initial sample of 260 participants because they did not report their gender. As a robustness test, we conducted all main analyses using the gender classification from Prolific Academic (Tables S2 and S3 in Multimedia Appendix 5). We used preintervention screening to ensure that none of the participants had participated in the first experiment and to balance gender representation. Table S1 in Multimedia Appendix 2 shows the summary statistics for the sample.

Participants were randomly assigned to 1 of 2 experimental conditions in which they viewed a short video conveying the same content about postdental care as that used in the first experiment. The conditions varied depending on whether the speaker was a female-presenting or male-presenting humanized AI avatar. Both humanized avatars were White, appeared to be of a similar age, and were dressed in casual clothes (the woman avatar wore a blouse and the man avatar wore a jacket over a T-shirt). The key outcomes examined were participants’ intention to comply and understanding of the video. The measures and analyses used to assess these outcomes, including the questions and scales, were identical to those in experiment 1.

This study evaluated AI-generated avatars for postoperative dental instructions. In experiment 1, the high level of realism of the humanized AI avatar was intentional, aiming to test whether near-human realism could achieve effects comparable to those of a real person. Results show that humanized AI matched a real speaker on intention to comply and comprehension, whereas animated avatars reduced intention to comply despite similar comprehension, suggesting that visual realism—not content—drives the observed responses. Regarding clinical significance, the small to medium effect size for the animated AI condition is comparable to effect sizes that have predicted real-world health behaviors in prior research [20]. Moreover, AI disclosure had mixed effects: it modestly lowered usefulness and engagement for humanized avatars but improved clarity and usefulness, although not engagement, for animated avatars. Transparency thus interacts with visual realism.

In experiment 2, women responded more positively to man avatars, while men did not exhibit a significant gender preference. This interaction highlights the complex dynamics between speaker characteristics and participant identity, indicating both opportunities and challenges for personalization strategies in AI-based health communication. Importantly, this interaction occurred without any difference in comprehension, suggesting that using a man avatar may enhance adherence among women without negatively affecting understanding for any participants.

Our finding that humanized AI avatars achieved functional equivalence with real speakers in terms of compliance intentions is consistent with the findings of Nasser Oesterreich et al [21], who found that photo-realistic avatars matched real photographs on trustworthiness ratings, and with those of Krieger et al [14], whose pilot study showed that virtual health assistants increased screening intentions. Our results suggest that this perceptual similarity extends to behavioral outcomes. Similarly, pilot studies by Haider et al [19] and Suárez et al [22] found that realistic AI avatars improved satisfaction and trust [19,22]; our study shows that comparable benefits persist in patient-facing instruction, although we measured intention to comply rather than attitudes toward the AI avatars. In addition, the compliance penalty we observed for animated avatars echoes the findings of Baake et al [10], who found that less realistic agents reduced trustworthiness in health communication. By holding informational content constant while varying the visual presentation (animated, humanized, or real person), our findings further demonstrate that these visual variations, rather than content clarity, drive compliance intentions.

Our findings on gender interactions align with broader patterns in avatar perception research. Eyssel and Hegel [12] found stereotype-driven preferences favoring man avatars on competence dimensions. We observed a similar pattern among women in a medical instruction context, while men showed no preference, suggesting that gender effects may be asymmetric in health care settings. Our disclosure findings show reduced engagement following AI revelation, parallel to prior observations following chatbot disclosure [18].

As AI tools become increasingly integrated into health care, it is crucial to examine how they might unintentionally reinforce social biases or stereotypes. Our findings show that women responded more favorably to man avatars. This may reflect internalized gender norms or stereotypes, potentially shaped early in life, that create expectations about professional roles and persist into adulthood [23]. This pattern is consistent with well-known social psychology theories, such as role congruity theory [15], which explains how stereotypes about ability can continue to exist.

Therefore, although personalization offers potential benefits, it must be ethically grounded. Designing avatars to match user preferences—whether in gender, race, or demeanor—should be done cautiously to avoid reinforcing stereotypes or limiting patient autonomy. Alternative approaches, such as user-driven selection of humanized AI avatars to deliver information, may enhance perceived autonomy; however, whether such approaches increase intentions to comply merits future investigation. These considerations point to the importance of using a population health framework to develop inclusive digital health interventions that enhance equity rather than exacerbate disparities.

This study contributes to the expanding literature on digital health communication and AI by empirically validating prior theoretical claims. Our results align with findings that human likeness and anthropomorphic cues increase trust and engagement [10,19]. They also offer nuanced insights into how AI disclosure can moderate user responses [24].

Furthermore, our findings support the view that comprehension of medical instructions can remain high even in AI-mediated formats, consistent with research on the role of animated computer agents in health literacy [7]. However, few studies have evaluated how these tools impact long-term adherence or real-world behaviors, which remains a critical avenue for future research.

The study was conducted in a single clinical context—post–dental implant care—and included a sample restricted to English-speaking participants in the United States, who identified as man or woman, which may limit the generalizability of the findings, particularly regarding gender interactions among nonbinary or gender-nonconforming individuals. Moreover, the observed effects could also reflect authority bias arising from attire cues (eg, a jacket), which might be confounded with gender effects, or from perceived professional role, as participants likely interpreted the speaker as a health care professional rather than as the dentists themselves who performed the procedure. Therefore, broader evaluations across multiple clinical settings and more diverse populations, including populations with substantially lower health literacy, are needed to assess the scalability and inclusiveness of AI-based communication tools. Additional measurement instruments may also be needed to detect improvements beyond the real-person baseline, which may have been subject to ceiling effects, as a benchmark to be reached. Additionally, experiment 2’s lack of a men human control group prevents assessment of whether women’s stronger preference for men avatars stems from an AI-specific positive bias or from more general health care gender biases, consistent with role congruity theory [15].

Future research should also examine how individual preferences, cultural background, and implicit biases influence user responses to AI avatars. Longitudinal studies that track actual health behaviors and outcomes will be essential to determine whether the observed effects on intention to comply translate into real-world adherence and improved health outcomes.

The findings suggest that humanized AI avatars could supplement or even replace human communicators in certain postcare contexts without compromising message effectiveness. Health care systems may leverage these tools to reduce clinician workload, provide repeatable instructions, and support patients with lower health literacy [7,25]. Designing avatars with clear voices, friendly tones, and culturally sensitive features—while strategically deciding when and how to disclose their AI nature—can maximize their impact [11,18,25]. If implemented ethically and equitably, these tools could help reduce postprocedure complications and lower health care costs, especially in resource-limited settings.

This study demonstrates that humanized AI avatars can be as effective as human providers in delivering postprocedure instructions, provided that their visual design and disclosure are thoughtfully managed. These findings contribute to the growing body of literature on AI in health communication and highlight the potential for personalized, scalable, and ethical digital care strategies. Moving forward, it is essential to ensure that such technologies are inclusive, empirically validated, and tailored to support equity in health care delivery.