What we build

Emotionally aware systems

Most AI systems process language with no account of the emotion behind it, or of the effect their own response will have. Brenkins builds systems that take both seriously. The idea has a clear lineage, and the evidence for why it matters is now arriving from the frontier of AI research.

A definition, written before the field caught up

In 2023, Jamie Brennan set out to define what a complete AI would need. Her thesis, A Study of Holistic Artificial Intelligence (HAI) and Gaps in the Concept, reviewed 49 definitions of artificial intelligence spanning a century and scored them against an established framework for concept quality. The gap she found was consistent: definitions described systems that compute, but not systems that understand the people they affect.

Her answer was Holistic AI:

The definition places Holistic AI within what the field calls Theory of Mind AI: a system that models the mental and emotional states of others and adjusts its behaviour accordingly. The thesis draws a careful line here. The system does not need to feel emotions. It needs to understand that the minds it interacts with do, and to treat that as a fact that shapes good decisions. In the thesis's terms, the goal is sapience, not sentience.

The same work named the limitation that still defines the debate. A language model, it argued, learns that the word "grief" tends to follow "death," but has "no understanding of the meaning of those words… beyond the statistical relationship between the tokens." Fluent emotional language is not the same as understanding what the language refers to.

From foresight to consensus

Brennan's argument was that real intelligence requires a model of the consequences of one's own actions and of the minds those actions affect. By 2026, that view had become a central thread in mainstream AI. Yann LeCun's case for world models makes the parallel claim from a different starting point: that genuine intelligence depends on a system that can anticipate the consequences of its actions, and that next-token prediction alone does not produce it. Brennan arrived there through emotional intelligence and theory of mind; LeCun through world models and robotics. They reached the same place independently, three years apart.

The convergence matters because it locates emotional awareness inside the larger question of machine understanding, and the harder question of machine consciousness: the difference between a system that generates convincing emotional text and one that registers that the text refers to real states in real people.

What the latest research shows

Until recently this was a conceptual argument. It is now also an empirical one. In a 2026 study, Anthropic's interpretability team mapped 171 distinct emotion concepts represented inside a single model (Claude Sonnet 4.5), and showed they are not surface vocabulary. Steering experiments demonstrate that turning one of these internal representations up or down produces predictable shifts in the model's behaviour, even when the input carries no emotional content at all. Emotion, in these systems, is structure in the computation, not decoration on the output.

Independent work points the same way. Research on where this processing happens (Tak et al., 2025) finds that emotion handling concentrates in the model's middle layers and resembles the cognitive-appraisal models psychologists use to describe human emotion. And on standardised tests of emotion identification (the SECEU benchmark; Wang et al., 2023), frontier models already score near or above human norms.

The honest gap

The standard model of emotional intelligence, from Salovey and Mayer, has four branches. Reading the current evidence against them shows where AI actually stands:

• Perceiving — identifying emotion in language and context Largely met
• Using — letting emotional information guide reasoning and attention Partial
• Understanding — reasoning about emotional blends, trajectories, and culture Partial
• Managing — regulating emotional state to serve a goal, without being told to Not yet

The sharpest gap is the fourth branch, and the sharpest point within it is autonomy. There is no evidence that current systems regulate their own internal emotional states on their own initiative rather than on explicit instruction. That is precisely why emotional awareness is an engineering and governance problem, not a solved feature.

How Brenkins approaches it

Emotionally aware does not mean emotional. It means a system that recognises the state of the person it is serving and weighs the effect of its response before giving it. We design for that, and we design for the limits above rather than around them: we treat the autonomy gap as real, keep behaviour verifiable rather than assumed, and stay close to the regulatory questions now forming around emotion-aware AI. The conceptual groundwork is a decade in the making. Our work is turning it into systems that can be built, measured, and trusted.

References

• Anthropic (2026). Emotion concepts and their function in a large language model. Anthropic. anthropic.com/research/emotion-concepts-function
• Brennan, J. L. (2023). A Study of Holistic Artificial Intelligence (HAI) and Gaps in the Concept. Master of Management in Business Innovation thesis, Bangkok University. dspace.bu.ac.th/jspui/handle/123456789/5750
• Cuzzolin, F., et al. (2020). Knowing me, knowing you: theory of mind in AI. Psychological Medicine, 50(7).
• Salovey, P., & Mayer, J. D. (1990). Emotional intelligence. Imagination, Cognition and Personality, 9(3).
• Tak, A. N., et al. (2025). Mechanistic interpretability of emotion inference in large language models. Findings of ACL 2025. aclanthology.org/2025.findings-acl.679
• Wang, X., et al. (2023). Emotional intelligence of large language models (SECEU benchmark). Journal of Pacific Rim Psychology. doi.org/10.1177/18344909231213958