Skip to content
Affective
← Research

Emotion is a surface pattern in every model you have used

2026-09-14

Historical note. This article records the September model thesis and may describe mechanisms or schedules that are no longer the active POC plan. Read the current status first.
Ratchet study: emotional context advances turn by turn and is held by the pawl, carried across turns rather than re-derived.PAWLT1T2T3T4T5T6SIDE ELEVATIONPERSISTENT ACROSS TURNSPRE-POC
Ratchet study. State carried across turns, rather than re-derived each time. Illustration.

Locally scoped, never persisted

Frontier models represent emotion as locally-scoped activations, re-derived per turn, never persisted. Interpretability work on frontier models shows emotion vectors track only the operative emotion needed to predict the next few tokens, then reset with each new context.

The consequences are structural: sycophancy, missed crisis signals in calm-sounding audio, emotional context forgotten across turns and sessions. None of this is fixable by adding parameters, because the problem is architectural.

The benchmark problem

Most claims of emotion recognition are confounded by lexical keyword detection. Remove emotion vocabulary from test items and performance collapses: the understanding was keyword matching, not comprehension.

EQ-Bench scores correlate with MMLU at r≈0.97. A model's emotional-intelligence score is mostly its general intelligence score, which is exactly what you expect when emotion is pattern-matched rather than represented.