Uneven Trust in AI: Experimental and Global Survey Evidence

Abstract

Hundreds of millions of people obtain information from the same AI models. Does their trust in AI depend on the technology or societal institutions? In an incentivized cross-country experiment, participants (N=2,928) trust identical GPT-4o answers differently, with a 0.57SD gap between the US and Kenya. Combining the experiment with three global surveys (N=546,985), we find that trust in AI is consistently linked with trust in traditional news institutions, and correlates with everyday AI attitudes. Leveraging this link, we provide exploratory predictions of global trust in AI, validated against independent surveys.