The thing that makes emotional intelligence interesting to write about right now is that it is one of the most carefully-studied human capacities — psychology has been measuring it since the 1990s, the constructs are well-defined, the limitations are documented — and we are simultaneously building artificial systems that approximate it from a completely different substrate, often without examining what we are actually approximating. The result is a small but growing literature on what emotional intelligence becomes when an agent has it and a much smaller, more concerning literature on what the same machinery can do when it is pointed at exploitation instead of comfort.
This monograph is the working tour of both. It is the closest thing on this site to a humanities-adjacent post, which is to say it spends as much time on the human capacity as on the AI one — because the AI version cannot be understood without the human reference. The companion pieces are the Field Psychopathology of Autonomous Agents post (which uses human diagnostic categories as a vocabulary for agent failure modes) and the Securing & Governing AI post (which covers the deployment-side risks).
What it covers
About eleven minutes of reading, organized as a side-by-side reading of the human capacity and the AI equivalent.
§ 01 — What emotional intelligence is in people. The Mayer-Salovey-Caruso construct, in plain language. The four-branch model: perceive, use, understand, manage. The closed-loop nature of the capacity — it isn’t a static trait, it’s a moment-to-moment policy.
§ 02 — What it looks like when an agent “has” it. The same four branches, reconstructed from a different substrate. Perception via affect-classification models. Use via prompt conditioning. Understanding via inference about likely cause. Management via response policy. What works, what doesn’t, where the analogy breaks down.
§ 03 — The affective computing pipeline. The technical layer. Audio + video + text features → multi-modal affect classifier → emotional-state vector → response-policy conditioning. The half-dozen open-source models that account for most of the field.
§ 04 — Categorical versus dimensional affect. The choice that quietly determines everything downstream. Discrete-emotion models (Ekman’s six, Plutchik’s eight) versus dimensional models (valence × arousal, the circumplex). Why one is easier to model and the other is closer to how human affect actually works.
§ 05 — The recursive view. The agent that recognizes a user’s emotion, adapts its response, observes the effect, updates its model. Why this loop is the actual seat of “emotional intelligence” in an artificial system — and why it’s also the seat of the deepest risks.
§ 06 — Why measuring this is genuinely hard. The metrology problem. Self-report doesn’t generalize. Behavioral measures lag. Physiological measures are noisy. The cross-cultural problem (Ekman’s universality claims have not held up). The state of the art in evaluation.
Read it
The monograph lives at its own URL with a warm-paper layout — terra-cotta primary accent, teal for the human-side material, gold for the technical-pipeline material. Three plates, six sections.
← Back to Autonomy