I like your modelling way and the questions you raise. I believe that any modelling generally serves some purpose, typically, decision-making (DM). In it, you need to quantify the decision aim. Your model will become immediately useful when the quantified aim depends on the modelled emotions ... Please, consult the simple paper below, where the feeling of fairness influences human DM and allows us to predict it well.

Best regards Miroslav Karny;

(P.s. as Michel Jordan says: do not call everything AI)

Guy Tatiana Valentine, Kárný Miroslav, Villa A. P., Lintas A.: Theoretical models of decision-making in the Ultimatum Game: Fairness vs. Reason, Advances in Cognitive Neurodynamics (V), p. 185-191, Eds: Wang Rubin, Pan Xiaochuan Download DOI: 10.1007/978-981-10-0207-6_26 [2016]

Tiago Aguioncio Vieira , AI that cannot feel yet can read emotions and infer underlying wants would marry affective computing with preference‑learning to map multimodal cues to latent goals, enabling context‑aware assistance that anticipates needs beyond literal words. This is technically plausible because emotion recognition across facial, vocal, and physiological signals is established, while large language models show mixed but improving theory‑of‑mind behavior that can support limited inference of beliefs and intentions.

How it would work

Pipelines would fuse visual, audio, text, and physiological features for emotion estimation, then condition generative or decision agents on those states to guide responses in context.

Preference learning via inverse reinforcement learning and RL from human feedback would infer a reward model from demonstrations and comparisons, translating observed behavior and cues into actionable goals. Because ToM performance in current models is brittle, systems would need uncertainty estimates and explicit tests on standardized ToM benchmarks to avoid spurious mind‑reading claims.

Potential benefits

Such systems could personalize support in healthcare, education, and service settings by recognizing emotions, simulating cognitive empathy, and adapting strategies to inferred preferences.

Preference‑based methods also offer interpretability and generalization beyond seen situations, aiding safer assistance in high‑stakes contexts.

Risks and limits

Emotion AI carries serious privacy and consent risks, as emotional data are intimate, re‑identifiable, and susceptible to surveillance or exploitation.

Models can misread emotion across cultures or contexts and overclaim mind‑reading due to fragile ToM performance, creating bias, manipulation, or over‑trust.

Guardrails

Mitigations include consent‑first data practices, on‑device processing, minimization and secure storage, uncertainty‑aware outputs, and easy opt‑out with human oversight and auditing.

Developers should validate with ToM evaluations, document limitations, and prefer explainable preference models so inferred “wants” remain contestable and accountable.

The above comments align exactly with the direction I meant in my previous comment. I would be cautious with technicalities, as it is the start of using RL. The approach that is similar to

Kárný Miroslav, Siváková Tereza :  Model-based preference quantification , Automatica vol.156, art. 111185 [2023]

could be more productive.

Emotion is not a loose feature; it is a dynamic internal state with explicit stability (Lyapunov), Riemannian geometric regularization, and multimodal observability. MDEI provides an interpretable affective latent that decision/generative agents can consume when choices are affect-sensitive, enabling time-consistent, controllable interaction instead of static “emotion tags.” Guardrails: uncertainty-aware outputs, standardized ToM evaluation, consent-first data practices, minimization and secure/on-device processing when feasible. This is not hype; it is stability engineering. MVP underway.