C5.01 - Oral Session: AI Avatars or Human Narrators? Testing COVID-19 Vaccine Messaging Outcomes

Brief Abstract Summary: Compare message outcomes between videos using AI-generated realistic avatars, AI animated characters, and human actors. Participants (N=1,001) were randomly assigned to one of three video conditions promoting COVID-19 vaccine uptake: (1) AI realistic human avatars, (2) AI animated cartoon-style avatars, or (3) human actors. Results showed that speaker rating was significantly higher for videos with human actors compared to AI cartoon-like animated videos but not compared to the AI hyper-realistic human avatar videos. There were no other differences in outcomes between conditions. Findings suggest that audiences respond less favorably to cartoon-style AI animated avatars than to human actors, while hyper-realistic AI avatars can perform comparably to human speakers. However, future research is needed to examine long-term effects.

Detailed abstract description:

Background: Generative artificial intelligence (AI) provides a new resource for public health to develop content quickly. Platforms such as ChatGPT and Synthesia can generate scripts, voices, and avatars ranging from cartoon-like animated characters to hyper-realistic human avatars. These tools have the potential to personalize messages for different communities. However, little is known about how AI-generated content compares to human-generated content in terms of audience perceptions. The Elaboration Likelihood Model (ELM) describes how audiences can use two routes to process information. The central route occurs when the audience is motivated for a personal reason (i.e., parent passed from COVID-19). For many individuals, they may not have that connection and instead use the peripheral route. The peripheral route occurs when they use a hueristics, such as perceptions towards the speaker or perceived credibility of source, to decide where to think critically about the information. If the audience can tell that video is made by AI and has negative perceptions towards AI, they may be less likely to accept its message. The purpose of this study was to compare message outcomes between videos using AI-generated realistic avatars, AI animated characters, and human actors.

Methods: A national sample of U.S. adults (N = 1,001) was recruited via an online survey panel. Participants were randomly assigned to one of three video conditions promoting COVID-19 vaccine uptake: (1) AI realistic human avatars, (2) AI animated cartoon-style avatars, or (3) human actors. Each participant viewed two videos (one male speaker, one female speaker) presented in random order. To reduce variation, each video included the same script and video length. Following each video exposure, participants rated a video on outcome measures including likeability, speaker rating, perceived effectiveness, perceived credibility, and intention to follow. All measures were taken from validated scales and reliability metrics were calculated before analysis. ANOVAs were used to compare outcome measures among the three conditions.

Results: Results showed that speaker rating was significantly higher for videos with human actors compared to AI cartoon-like animated videos but not compared to the AI hyper-realistic human avatar videos. There were no other differences in outcomes between conditions.

Discussion: Findings suggest that audiences respond less favorably to cartoon-style AI animated avatars than to human actors, while hyper-realistic AI avatars can perform comparably to human speakers. This may suggest that using hyper-realistic AI software, such as Synthesia, may produce similar message effects. However, limitations to the study include testing only a single exposure and did not assess longer-term effects. Future research should explore cost-effectiveness and long-term effects of using generative AI for health messaging, as well if same effects would be seen among children.