Evaluating the Credibility of Human and AI Forecasters: Receivers Discriminate Performance More Finely for Humans

Le résumé fourni par la source

Decision-makers in high-stakes forecasting domains are increasingly asked to weigh probabilistic judgments from human experts alongside those from artificial intelligence (AI) systems. Whether source attribution alters how forecasts are received has been largely unexamined. Three accounts make competing predictions. Algorithm aversion and algorithm appreciation predict a source dependent bias in favor of humans and AI, respectively: both assume that source attribution may affect forecaster evaluation. By contrast, responsibility-attribution research predicts differential discrimination, with credibility ratings distinguishing performance levels more finely for human than AI sources. We tested these accounts in two experiments (N1 = 667; N2 = 1,170). Participants translated verbal probability expressions communicated by either a human expert or an AI system into numeric point estimates and ranges. Separately, participants evaluated source credibility after reviewing a forecaster’s performance record, with accuracy manipulated between subjects. Interpretation was largely unaffected by source: numerical point estimates were similar and ranges equivalent for both human and AI sourced forecasts. In contrast, credibility ratings depended on source: receivers discriminated performance more finely for human than AI forecasters: human forecasters were rated as more credible than equally accurate AI when accuracy was high and less credible when accuracy was low. This differential discrimination is inconsistent with algorithm aversion and appreciation accounts, but consistent with the responsibility-attribution research, according to which people hold human advisors more accountable for outcomes than algorithms. We discuss implications for integrating AI-assisted forecasts into decision support roles.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.