Balancing realism and trust: AI avatars in science communication

In plain English

An experiment with 491 German participants tested whether a more realistic AI avatar explaining a science topic is trusted less — the uncanny valley prediction. The opposite happened: higher realism raised ratings on all three trust dimensions the study measured (expertise, integrity, benevolence). Male avatars were rated higher on expertise only, and a viewer's prior AI knowledge and trust in science mattered as much as the avatar itself. The caveat practitioners should carry: one video per condition, one topic, one country, so this is a reason to stop assuming realism is a liability rather than proof that it is an asset.