The media has portrayed certain artificial intelligence (AI) software as committing moral violations such as the AI judge of a human beauty contest being “racist” when it selected predominately ...light-skinned winners. We examine people's attributions of morality for seven such real-world events that were first publicized in the media, experimentally manipulating the occurrence of a violation and the inclusion of information about the AI's algorithm. Both the presence of the moral violation and the information about the AI's algorithm increase participant's reporting of a moral violation occurring in the event. However, even in the violation outcome conditions only 43.5 percent of the participants reported that they were sure that a moral violation occurred. Addressing whether the AI is blamed for the moral violation we found that people attributed increased wrongness to the AI – but not to the organization, programmer, or users – after a moral violation. In addition to moral wrongness, the AI was attributed moderate levels of awareness, intentionality, justification, and responsibility for the violation outcome. Finally, the inclusion of the algorithm information marginally increased perceptions of the AI having mind, and perceived mind was positively related to attributions of intentionality and wrongness to the AI.
•AI's real-world moral violations only reported 43.5% of the time.•Knowledge of algorithm increases moral violation perception.•AI not organization, programmer, or use is attributed wrongness.•AI attributed mind, awareness, intentionality, justification, and responsibility.
Artificial intelligence (AI) agents make decisions that affect individuals and society which can produce outcomes traditionally considered moral violations if performed by humans. Do people attribute ...the same moral permissibility and fault to AIs and humans when each produces the same moral violation outcome? Additionally, how do people attribute morality when the AI and human are jointly making the decision which produces that violation? We investigate these questions with an experiment that manipulates written descriptions of four real-world scenarios where, originally, a violation outcome was produced by an AI. Our decision-making structures include individual decision-making - either AIs or humans - and joint decision-making - either humans monitoring AIs or AIs recommending to humans. We find that the decision-making structure has little effect on morally faulting AIs, but that humans who monitor AIs are faulted less than solo humans and humans receiving recommendations. Furthermore, people attribute more permission and less fault to AIs compared to humans for the violation in both joint decision-making structures. The blame for joint AI-human wrongdoing suggests the potential for strategic scapegoating of AIs for human moral failings and the need for future research on AI-human teams.
