The Anchoring Effect, Algorithmic Fairness, and the Limits of Information Transparency for Emotion Artificial Intelligence.

This article investigates the performance and fairness of emotion artificial intelligence (AI) systems in recognizing emotions from facial expressions, focusing on how human labelers interact with AI outputs under conditions of information transparency about algorithmic fairness. The study uses three commercially available emotion AI tools—Face++, Microsoft Azure Face API, and Amazon Rekognition—and two facial image data sets (NBA players and Balanced Faces in the Wild) to examine inconsistencies in emotion labeling and the influence of AI scores on human judgments, highlighting the anchoring effect where humans rely on AI outputs even when aware of potential biases. Key findings include significant variability and racial disparities among emotion AI models, a stronger anchoring effect in more difficult emotion recognition tasks, and limited or mixed impact of transparency about AI fairness on improving human labeling accuracy; in some cases, transparency increased human inconsistencies. The research underscores the challenges of mitigating algorithmic bias through individual decision-making and suggests that transparency alone is insufficient to ensure fair outcomes, emphasizing the need for comprehensive policies and computational debiasing in emotion AI applications.