Research
Affective AI and Business
We re-frame the meeting point of uncertainty, acceptance, and ethics in industrial deployment as practitioner knowledge.
What we work on
When affective AI is built into corporate activity, technical accuracy alone is not enough in scene after scene. User acceptability, operational uncertainty, accountability, and the ethical considerations specific to services that touch emotion — these multiple axes of constraint must be satisfied simultaneously. This domain takes as its object the design of affective AI not for 'selling' but for 'delivering while protecting people,' and explores diverse applications including recommendation, content moderation, and decision support.
What we are finding
Studies of multimodal recommendation and consumer protection have made the essentials of implementing affective AI visible: designs that extract latent orientations such as 'lifestyle,' techniques that adjust the affective intensity of stimulating information while preserving meaning, and longitudinal monitoring of use context. These are effective for simultaneously satisfying technical accuracy and social acceptability.
Research notes
Research narrative
When affective AI is built into corporate activity, technical accuracy alone is not enough in scene after scene. A recommender system reads the user's emotion too well and invades privacy; content delivery competes on stimulation and worsens users' mental state; customer-service AI accumulates personal emotional information and leaks it to third parties — all are problems that surface only in deployment. This domain takes as its object the design of affective AI not for 'selling' but for 'delivering while protecting people,' and explores diverse applications including recommendation, content moderation, and decision support.
Industrial deployment imposes constraints absent in research prototypes, all at once. First, user acceptability: even technically excellent systems will not spread if users find them 'creepy.' Second, operational uncertainty: in real environments, inputs unlike training data flow in daily, and model behavior becomes unpredictable. Third, accountability: when something goes wrong, commercial deployment requires clarity about who is responsible. Fourth, ethical considerations specific to emotion-related services: because they intervene in users' vulnerable moments, they demand a higher ethical bar than ordinary IT services. This domain advances design research that handles these constraints simultaneously.
Commercial use of emotional data carries distinctive ethical problems. First, although emotional data is sensitive personal information, in many cases it is not explicitly covered by current law. Second, the accuracy of emotion prediction is also the accuracy of manipulating users, and it can distort decision-making through micro-targeted advertising and the like. Third, it is practically difficult for users to refuse the acquisition of emotional data entirely (webcams, microphones, and text are everywhere). This domain explores technical and normative designs that respond to these problems together with concrete applications.
MALLET (Multi-Agent LLM-based Emotion Tempering, 2026) is a consumer-protection framework that delivers stimulating news and similar expression to users with reduced affective intensity while preserving meaning. The background is the concern that the 'attention economy' of social media and news distribution captures attention by stimulating people's emotion, ultimately worsening users' mental state. MALLET consists of four LLM agents. An emotion-analysis agent measures the affective intensity of the input text, a tempering agent rewrites the text to lower its affective intensity while preserving meaning, a monitoring agent tracks the user's emotion history weekly, and a guide agent produces individualized feedback. On 800 AG News items, MALLET achieves stimulation-score reductions of up to 19.3% while keeping SBERT similarity at 0.83 or above, jointly preserving meaning and tempering emotion.
GNN-Enhanced Multimodal Fusion (2025) rethinks user personalization through the concrete application of meal recommendation. Conventional meal recommenders are dominated by collaborative filtering that recommends similar recipes from past recipe ratings, but this ignores the user's long-term health goals and lifestyle. This work integrates diverse information — food images, nutritional data, ingredient composition, and cooking methods — using a Graph Neural Network (GNN). A GNN is a deep-learning method that performs representation learning while taking relations between nodes into account, and it can consistently handle diverse relations such as ingredient-to-dish and dish-to-nutrition. Furthermore, contrastive learning extracts the user's 'lifestyle' as a latent representation. As a result, the model surpasses existing methods on AllRecipes data and demonstrates the feasibility of food recommendations consistent with health goals.
Both studies share the orientation of designing AI not for 'selling' but for 'delivering while protecting people.' MALLET deliberately reduces affective intensity to prioritize long-term user wellbeing over short-term attention maximization. GNN-Enhanced Multimodal Fusion prioritizes long-term lifestyle alignment over short-term preference matching. Both are studies that articulate the 'point where accuracy and ethics balance' in industrial deployment as concrete design choices. This domain will concretize, application by application, this 'people-centered way of deploying affective AI.'