Technical Solutions to Emotion AI’s Privacy Harms: A Systematic Literature Review

Emotion AI, while contested for its validity, bias, and accuracy, claims the ability to infer individuals’ emotions and other affective qualities. While proponents tout its potential (e.g., improving well-being and productivity), critics raise concerns about its impact on data subjects’ privacy. We conducted a systematic literature review of scholarship that has explored technical solutions to address emotion AI’s privacy concerns, examining the underlying conceptualizations of data subjects (i.e., individuals subjected to emotion AI), data, and privacy that motivated them. Findings reveal patterns of 1) conceptualizations of data subjects as decontextualized and flattened, 2) a heavy focus on the sensitivity of input data for emotion AI systems while neglecting the sensitivity of output data (i.e., emotion inferences), 3) conflating privacy with security and 4) viewing privacy as a burden to the development of large-scale emotion AI systems. We argue these conceptualizations motivate technical solutions which largely fail to address the full range of emotion AI’s privacy harms. We discuss what a human-centered and comprehensive conception of privacy would mean for emotion AI development, concluding that while technical approaches can address some privacy concerns, key privacy concerns persist.

Download paper here