Distinguishing Emotion AI: Factors Shaping Perceptions Including Input Data, Emotion Data Recipients, and Identity

Emotion artificial intelligence (emotion AI), which claims to infer data subjects’ emotions, is used across a range of high-stakes con- texts in the U.S. Examining data subjects’ perceptions of emotion AI is important given the impact its use can have on their access to opportunities, among other consequences. Such inquiry requires an empirical understanding of data subjects’ attitudes about emotion AI compared to AI that does not infer emotions, and differences across potential 1) input data (e.g., facial data, speech data), and 2) recipients of its outputs (i.e., emotion data). To these ends, us- ing a contextual integrity lens, we conducted a survey (n = 742) with a representative U.S. sample, oversampling for individuals of minority groups. Results demonstrate data subjects’ attitudes toward emotion AI are significantly more negative than AI that does not detect emotion. Further, we find significant contrasts in perceptions of specific input data across gender, race, and disabil- ity status, particularly for neurodiverse individuals. We also find participants express concern with all emotion data recipients, with higher concern for their data being accessed by recipients with large power asymmetries. Based on these findings, we argue emotion AI should not be deployed without a comprehensive assessment of its implications across demographic groups, yielding implications for policymakers and developers.