Date created: | Last Updated:

Category: Project

Description: ABSTRACT. Advances in artificial intelligence (AI) raise important questions about whether people view the moral evaluations of AI systems similarly to human moral evaluations. We conducted the first moral Turing Test by asking people to distinguish real human moral evaluations from those of a popular advanced AI language model. A representative sample of 299 U.S. adults first rated the quality of moral evaluations when blinded to their source. Surprisingly, they rated the AI's moral reasoning as superior to humans’ along almost all dimensions, including virtuousness, intelligence, and trustworthiness. Next, when tasked with identifying the source of each evaluation (human or computer), people performed significantly above chance levels. Analysis of follow-up questions suggests that multiple factors could have cued participants to the true source of the computer evaluations, including their greater perceived quality and some structural differences. Although the AI failed our moral Turing Test, ironically, this was not because of its inferior moral reasoning but, potentially, its perceived superiority. This perception raises concerns that people may uncritically accept an AI’s potentially harmful moral advice, highlighting the need for safeguards around AI language models on matters of morality. NOTE: See Files section for study data, coding labels, and stimuli.