Ethics and Philosophy of Affective AI

AI that handles emotion is no longer a mere classifier; it is a socio-technical system in which heterogeneous layers — psychological theory, emotion-labeled data, machine learning models, dialogue interfaces, social applications, and ethical responsibility — are stacked on top of each other. To let a computer handle inner experiences such as joy and sadness, one must first choose a psychological theory of what emotion is, then choose a labeling scheme grounded in that theory, then design a model that learns those labels, then design an interface that presents results to the user, and finally articulate the ethical norms required for real-world deployment. These choices form a connected chain. Research in this domain consciously surfaces this layered structure and addresses the normative problems lurking at its seams.

Within the affective-AI research community, each layer has developed its own methodology and evaluation criteria in relative isolation. Psychologists refine emotion models, data scientists refine annotation schemes, machine learning researchers compete on benchmark accuracy, HCI researchers refine dialogue design, applied researchers push industrial deployment, and ethicists discuss norms. As each specialty deepens, a 'vacuum of responsibility' opens up between layers: problems that should be examined as connected get sliced apart at layer boundaries. This domain treats that pattern of disconnection itself as an object of study.

Bridging the Silos in Affective AI (2026) is a position paper that organizes affective-AI research as a six-layer pipeline (theory, data, model, dialogue, social application, ethics/evaluation) and diagnoses four recurring patterns of disconnection — 'silo bridges' — between layers. Specifically, it identifies (i) operationalization drift between theory and data (the emotion that theory speaks of and the label that ends up on data diverge), (ii) cognitive mismatch between model and dialogue (model outputs and the user's interpretive frame fail to mesh), (iii) the dissipation of responsibility between technology and ethics (it becomes unclear who is ultimately accountable), and (iv) implicit representativeness assumptions between data and social application (data from specific populations is treated as universal truth).

Why does this require a position paper? Because these disconnections are not isolated technical bugs but problems rooted in the very structure of the research community. A model that achieves state-of-the-art accuracy on a particular benchmark may still mislead users at deployment time, and improving the model alone will not fix that. The evaluation axes of the field themselves must be questioned, and a common language that crosses layers must be proposed. This paper aims to put such a cross-cutting perspective before the academic community.

As a response, the paper proposes five interlocking Design Criteria (DC). DC1, 'theory disclosure,' requires that the adopted emotion theory be explicitly stated in both papers and implementations. DC2, 'intervention boundaries,' requires that the scope of legitimate intervention by affective AI be defined in advance and that the design prevent overreach. DC3, 'longitudinal evaluation,' makes it essential to track effects over time rather than measure accuracy at a single moment. DC4, 'deployment-specific accountability,' requires that — across research prototypes, field trials, and commercial deployment — it be made explicit who bears responsibility for what. DC5, 'preservation of users' interpretive authority,' bakes into the design the principle that the final right to interpret an AI output belongs to the user. These five are not independent rules but a mutually reinforcing normative system.

DC5 — preservation of users' interpretive authority — sits at the core of the 'emotional sovereignty' concept that this domain puts forward. Emotional sovereignty is the normative position that the final right to interpret, record, and disclose one's own emotions belongs to the person experiencing them. It demands that the asymmetric power relation between an AI saying 'you are angry' and a person replying 'I do not feel that way' be consciously dissolved at the design stage. The concept is positioned as an extension of the right to self-determination in medicine and the right to informational self-control in data protection, and is proposed as a new human-rights notion for the age of affective AI.

The social and academic reasons that philosophical and normative work on affective AI is needed right now are clear. First, the rapid spread of generative AI is pouring systems that read and respond to emotion into everyday life, and treating their behavior as a problem only after the fact is no longer fast enough. Second, even though emotional data is sensitive personal information, large areas of it are not adequately covered by current law. Third, a shared language among engineers, users, and regulators has yet to mature, and discussion tends to split between technical and ethical registers. Research in this domain aims to present a framework that fills these three gaps.

This position paper is not a technical paper but a map for re-situating affective AI as a socio-technical system, and it functions as a reference frame for subsequent implementation research and ethical debate. The other domains of our lab — augmenting emotional data, understanding the inside of affective AI, human–AI interaction, business application, and psychological-support development — can each be understood as concrete research carrying one of the layers in this map. Ethics and philosophy serve as the warp thread that runs through all the other research.