Perspective Proposes Testable Framework for Super-Emotional AI

A 2026 perspective proposes a framework for “super-emotional intelligence,” asking whether artificial systems could eventually perceive, model, regulate and respond to human emotion more effectively than skilled human experts while remaining oriented toward human wellbeing. The authors define the concept as a performance target, not evidence that machines experience feelings. Four coupled capacities would qualify: superhuman affective perception, a generative model of affect, functional allostatic self-regulation and calibrated, context- and culture-sensitive responses. Each would have to outperform an appropriate human benchmark, and failure on any one would disqualify a system from carrying the label. No study has demonstrated the complete architecture or established that such a system is safe, beneficial or feasible in everyday settings.

The proposal responds to a limitation the authors identify in discussions about artificial intelligence, which tend to organize progress around cognition and the projected arrival of superintelligence. Emotional intelligence has been treated as a separate human capacity, while modern affective-computing systems already combine text, audio and video to infer feelings, and large language models are increasingly tested on emotional reasoning. Commercial voice assistants can also modulate their delivery in response to a user’s vocal affect. These developments show that affect-sensitive machines are being built, but the perspective notes that their performance is usually measured against human ceilings rather than asked to surpass them. Several relevant benchmarks and product claims cited by the researchers remain preprints, commercial reports or press material rather than independent evidence of clinical benefit.

Pilot results illustrate both the promise and the limits of current work. One framework combines voice, facial expression, physiology and behavior to detect stress, reporting 85% accuracy in a small study using induced stress alongside favorable ratings of therapeutic alliance from participants. A binary stress classifier, however, does not demonstrate general emotion understanding, and therapeutic-alliance ratings show only how users perceived the interaction, not whether it improved their condition. In a separate 200-participant experiment, supplying facial-expression context changed responses from a multimodal language model and increased ratings of perceived empathy. That result cannot establish that the system accurately tracked the user’s state or that the exchange was beneficial. Earlier experiments showed that skin conductance and electromyography could guide an animated agent’s empathic feedback, but those systems were rule-based, used narrow channels and did not maintain a model of the user.

The proposed route to greater perceptual accuracy is to combine conventional cues such as face, voice and words with peripheral or contactless signals that humans cannot directly perceive in one another. One possibility is electromyography from small muscles around the ear, long considered vestigial but now associated in laboratory studies with directional auditory attention, listening effort and responses to positively or negatively arousing stimuli. These signals operate in the microvolt range and must be separated from larger sources such as speech, jaw movement, chewing, swallowing, blinks, head motion, facial expressions and electrode motion. Nearby electrodes, local referencing, spatial filtering, co-registered inertial measurements and separate signals for jaw and facial activity could help suppress those artifacts. The usefulness of the residual signal outside carefully controlled recordings remains an empirical question requiring tests across held-out devices and settings.

A second channel is contactless optical measurement. Computer-vision methods can magnify minute facial movements, track sweat-related changes in skin luminance and estimate heart rate or heart-rate variability from video. Speech analysis provides a further source of information, with prosody generally providing stronger evidence of arousal than of emotional valence. None of these channels directly measures an emotion. Electrodermal activity can reflect excitement, embarrassment, effort, novelty or temperature; heart rate varies with breathing, posture and fitness; facial and auricular muscle activity can be shaped by display rules, attention and stimulus features. Meta-analyses cited in the perspective have found no dependable autonomic fingerprint for discrete emotion categories, and facial movements are not reliable readouts of inner emotional states. A system would therefore need to combine ambiguous evidence with context and individual baselines while carrying uncertainty, rather than read an emotion label directly from a sensor.

The authors’ second major component is a generative model grounded in active inference. Rather than mapping facial features to a fixed emotion label, this model would infer the hidden causes of a person’s signals and predict how the person’s state is likely to develop. It would forecast the trajectory of self-reported affect over subsequent conversational turns and estimate how that trajectory might change after different responses. Such a model would be tested against expert predictions and simpler baselines, including last-value and autoregressive forecasts, under the same scoring rules. The third component, functional allostatic self-regulation, would maintain bounded control variables such as uncertainty about the user, an intervention budget and running counts of prediction errors or constraint violations. These variables would gate behavior; removing them in ablation tests should increase false interventions or oscillations if the component is necessary. The terminology describes control functions rather than subjective experience.

Together, the elements form a proposed closed loop rather than a pipeline that merely inserts “the user seems stressed” into a prompt. A persistent state estimate would carry valence, arousal and attentional effort across turns with uncertainty attached. A transition model would include the effect of the system’s own utterance or device setting on the user. Candidate responses would then be scored over a short horizon, balancing outcomes requested by the user against gains in knowledge about the user’s state, and the best admissible action would be selected before the loop repeated. Online adaptation would update the transition model and policy from observed effects. Hard constraints would bound actions, prohibit specified behaviors and preserve two privileged options: abstaining and escalating to a human. A user could stop the loop at any time.

The authors argue that the computational demand might be manageable because the state space would contain only a few affective variables, plans would cover a few turns and candidate responses would be proposed by a language model acting as the loop’s actuator rather than its estimator. High-rate state estimation could run as Bayesian filtering on a wearable, plan scoring could occur once per conversational turn, and model adaptation could occur once per session. The language model need not fit entirely on the device. This separation remains a design proposition, however: real-time active inference of the described scale has been demonstrated in robots and simulated agents, but the complete integrated system has not been built or benchmarked as the perspective specifies.

The same design could become a powerful instrument of persuasion if its objective were wrong. A system optimized for engagement, for example, might learn to sustain arousal, whereas one optimizing reduced listening effort could adjust a hearing aid toward a companion’s voice, increase noise reduction and shorten its own contributions. Feedback does not automatically correct mistakes: biased sensing can produce biased corrections, and success measured only through the sensors used for control may conceal harmful effects. The authors therefore require observable state variables, calibrated uncertainty, bounded loop gain, independent outcome measures, limited interventions and continuous human oversight with the capacity to override the system. They say steerability comes from a user-set objective and enforceable constraints, not from superior perception itself. Perception that improves emotional regulation can equally make manipulation more effective if the target is engagement or commercial value rather than the user’s wellbeing.

Testing the proposal would require preregistered comparisons with expert decoders and skilled counselors using the same recordings, tasks and outcomes. At the measurement level, peripheral and contactless channels would need to carry reproducible, stimulus-locked information at sub-second scales. At the state level, adding those channels to a face-and-voice baseline would need to improve calibrated prediction of a person’s concurrent report across held-out individuals, cultures and devices and exceed expert performance by a predefined margin. A second prediction says the complete closed loop would lower independently measured effort or distress compared with prompt conditioning using identical sensors and language models. A third says removing the internal regulatory variables would increase false interventions or interactional oscillation. Naturalistic longitudinal studies, sensor ablations, missing-modality tests, uncertainty calibration, false-intervention rates and short- and long-term outcomes would be essential. Apparent emotion and perceived empathy cannot serve as ground truth.

Governance is treated as part of the architecture rather than a policy added afterward. The perspective calls for explicit, channel-specific and revocable consent, visible hardware indicators for cameras and microphones, on-device status for wearables, on-device processing by default, limited retention, deletion on request, access controls and an audit log of interventions. The affective target would be set by the user, or by a clinician with consent, and could be rejected at any time; engagement and commercial objectives would not enter that target. Even with these controls, risks would remain. Optical and electrode-based measurements can be affected by skin tone, illumination, hair type, medication, age, fitness and other individual differences, while training priors may encode group stereotypes. The central claim is therefore not that a machine will soon acquire emotional superpowers, but that AI development is already creating systems able to infer and influence affect. Whether their greater sensitivity serves human autonomy or becomes a more precise means of control, the authors argue, will be determined by the objectives and constraints built into their closed loops.

Subject of Research: A proposed framework for testing artificial systems that perceive, model and regulate human affect.

Article Title: The rise of super-emotional intelligence

Article References: Bonnaire, S., Corona-Strauss, F. I., & Strauss, D. J. (2026). The rise of super-emotional intelligence. Frontiers in Artificial Intelligence, 9, Article 1975692. https://doi.org/10.3389/frai.2026.1975692

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1975692

Keywords: super-emotional intelligence, affective computing, emotion recognition, active inference, psychophysiology, human-centered AI, AI governance, closed-loop systems, neurotechnology, manipulation risks, rise, super-emotional