The Artificial Intelligence Disclosure Penalty: Humans Persistently Devalue AI-Generated Creative Writing

Abstract

Although preliminary evidence suggests that humans often react aversely to artificial intelligence (AI)-generated creative works, we have little understanding of how robust or persistent these reactions may be. In a series of 16 preregistered experiments (N = 27,491), we examine how evaluations of creative writing are affected by whether participants believe the content is produced with an AI model. We find consistent evidence of an AI disclosure penalty: Participant evaluations of creative writing decrease when they believe writing samples were written by an AI model—or with the help of one—rather than a human author alone, and this effect is mediated by perceived authenticity. The AI disclosure penalty is sticky, persisting across evaluation metrics, contexts, kinds of written content, and multiple interventions derived from prior research aimed at moderating the effect. Our results indicate that AI disclosure penalties about creative writing may be stubbornly difficult to mitigate, at least at this time.