Bridging the Silos in Affective AI: A Critical Perspective from Data to Society

Affective computing has grown into a field that spans emotion theory, multimodal data, computational modeling, human-AI interaction, social applications, and ethical evaluation, yet most existing surveys address only one or two of these layers at a time and fail to capture how the layers depend on one another. Rather than offering another comprehensive survey, this position paper reorganizes the field around a six-layer pipeline and reads the literature as a chain of dependencies in which upstream design choices propagate into downstream evaluation, deployment, and governance. Our central diagnosis is that four silo-bridge patterns recur across the field: disconnections at the theory/data (Pattern~1), modeling/interaction (Pattern~2), technology/ethics (Pattern~3), and data/social-application (Pattern~4) boundaries. These are not incidental implementation problems but recurring failure modes rooted in the pipeline structure itself, forming a cascade in which upstream patterns structurally enable downstream ones. As a prescriptive program, we present five integrated design criteria: theory-explicit affective representation (DC1), affective reasoning with bounded intervention (DC2), longitudinal interaction-in-the-loop evaluation (DC3), deployment-specific affective accountability (DC4), and user-retained interpretive authority (DC5). The five criteria are designed as an interlocking response to the cascade rather than as independent recommendations. We situate affective AI not as a mere recognition or generation technology but as a sociotechnical system that can intervene in human emotional interpretation, interaction, and social decision-making, and call upon future research to move toward systems that are more coherent, more responsible, and more respectful of human emotional agency.