Turn a Blind AI: The Impact of Compassion Fade and the Identifiable Victim Effect on AI Risk Concerns

Abstract

Rapid recent advancements in artificial intelligence (AI) have prompted concerns from experts about its long-term implications for society. It is unclear how best to communicate with the general public about AI risks, particularly when understanding and acknowledgment is low in comparison to experts. Cognitive biases may impact risk perception, making them an important area to investigate when it comes to improving risk communication. This online experiment explored how vignettes incorporating the identifiable victim effect influenced concern about AI risks. We found that vignettes portraying identifiable victims elicited stronger concerns than those with statistical victims. The effectiveness of the identifiable victim effect remained consistent across different types of risks (e.g., biased criminal justice decisions, deep fake pornography). Notably, the increase in concern was found to be mediated by the vividness of the narratives, but not by the perceived controllability of the risks. Our findings offer valuable insights into AI risk communication, suggesting that narratives with identifiable victims are more likely to increase public concern.

Keywords: Compassion fade, identifiable victim effect, scope neglect, risk perception, artificial intelligence

JEL Classification: O33