Generative AI Are More Truth-Biased Than Humans: A Replication and Extension of Core Truth-Default Theory Principles

Abstract

Humans often display a truth-bias—the perception that others are honest independent of message veracity—but does this phenomenon extend to generative artificial intelligence (AI)? We had humans and large language models make nearly 1,000 veracity judgments across different prompts. Human detection accuracies were near chance (50%–53%) with notable truth-biases (59%–64%); AI had a substantially greater truth-bias than humans (67%–99%). GPT-4 was also truth-default, not suspecting deception when veracity assessments were unprompted. Together, people and AI judge most information to be true.