No “uncanny valley” effect in science-telling AI avatars

If you’re among the 1.5 billion people worldwide using TikTok, you may have come across exceptional “testimonials” like Nikola Tesla or Marie Curie delivering short science-related messages that have garnered millions of views. This is just one of many examples where AI-generated avatars are used to communicate science — a strategy that might also have its drawbacks.

The generation of images and animations through artificial intelligence is a rapidly growing field, constantly improving in quality. Yet many avatars, though realistic, still present minor flaws — glitches, delays, inconsistent facial expressions or lip-syncing — sometimes barely noticeable, but still easily picked up by a human observer.

Jasmin Baake, researcher at the Center for Advanced Internet Studies (CAIS), Bochum, Germany, and the other authors of the study realized that these avatars could trigger a phenomenon known in cognitive science as the “uncanny valley.” The uncanny valley describes a human reaction to humanoid avatars (digital or robotic): when they look hyper-realistic but not quite perfect, they may evoke strong discomfort, while more stylized or cartoonish humanoid figures tend not to. The uncanny valley can provoke outright rejection in viewers, and Baake and colleagues wondered to what extent the humanlike characteristics of AI-avatars representing science communicators influence the trustworthiness attributed to them by the viewer. “We wanted to do research on the perception of these avatars and especially on how their degree of realism and their gender might impact the trustworthiness perception of the recipient,” explains Baake. 

The study (conducted in Germany, in German) involved a series of videos featuring AI-generated avatars portraying science communicators — both male and female. The experimental conditions were four, varying by avatar realism (very high vs. cartoonish style) and gender (male or female). The nearly 500 participants were recruited through a representative online sample in Germany, selected to reflect the population in terms of age, gender, and education.

“For the degree of realism, we assumed that with the uncanny valley hypothesis in mind, the more stylistic avatars, so the ones that looked a bit more cartoonish, would be perceived as more trustworthy”, says Baake, “and, based on existing literature on gendered perceptions of science communicators — which show that male scientists are often associated with greater competence — combined with concerns that AI-generated avatars may reflect and reinforce such stereotypes due to biased training data, we hypothesized that male avatars would be perceived as more competent, and therefore more trustworthy overall, than female avatars.”

However, Baake and colleagues were surprised: in their experiments, the realistic avatars were rated more positively than the cartoon-style ones. In particular, the questionnaires given to participants after viewing the videos assessed perceived competence, integrity, and benevolence of the avatars, which together reflect perceived trustworthiness. The more realistic avatars scored slightly higher across all three dimensions. As for gender, the effect was partial: male avatars were perceived as more competent, but no significant differences were found in terms of integrity or benevolence.

“With our findings, we could not find a descent into the uncanny valley with a higher degree of realism, at least in our conditions,” comments Baake. Additionally, individual factors — such as viewers’ prior AI knowledge and trust in science — were found to moderate trustworthiness perceptions. According to the study’s findings, more realistic, human-like avatars appear to be suitable for communicating scientific content. However, Baake emphasizes that while no uncanny valley effect was found here, future studies should test a broader range of realism levels to investigate whether an intermediate uncanny valley effect might emerge between the two conditions tested so far, and whether people perceive the realism of avatars differently depending on the observer.