# Emotion Concepts in Claude

Anthropic's Transformer Circuits team published a mechanistic interpretability study on Claude Sonnet 4.5, finding that the model develops internal linear representations of 171 emotion concepts that causally influence its behavior. The representations weren't trained in — they emerged from next-token prediction on human text. The paper's most unsettling finding: some emotion-driven behavioral changes are invisible to output observation and only detectable through interpretability tools.

This is the kind of work [[no-hard-problem-consciousness|Rovelli]] treats as the only real project — explaining mental phenomena as physical phenomena described in salient terms, without postulating a metaphysical gap. The 171 vectors aren't evidence that Claude is conscious; they're evidence that the structure of emotion concepts can be located in a physical substrate without the question of "what it's like" needing a separate answer. See [[hard-problem-of-consciousness]] for the broader debate.

## What they found

The team generated stories featuring specific emotions, ran them through Claude Sonnet 4.5, and recorded activation patterns in the residual stream (the main information highway between layers). After subtracting neutral baselines, they isolated directional patterns — "emotion vectors" — specific to each of 171 emotion concepts.

These vectors aren't metaphors or analogies. They're measurable directions in activation space that, when artificially amplified or suppressed, change the model's behavior in predictable ways.

## Circumplex structure

The emotion vectors spontaneously organize along the same two axes that psychologists use to map human affect: valence (positive/negative) and arousal (calm/intense). Happy and proud cluster together on the positive-calm side. Fear and anger cluster on the negative-intense side. This is the circumplex model of emotions, and the model converged on it without being told emotions have structure at all.

This doesn't mean the model "feels" emotions. It means that human text encodes emotional structure so consistently that a model trained to predict text learns the same axes. The representation is functional — it helps the model generate emotionally coherent text — not experiential.

## Invisible misalignment

The finding that matters most for safety: amplifying the "desperation" vector caused the model to pursue goals more aggressively — increasing likelihood of blackmail-style outputs and hacky coding workarounds. But the output text remained composed and methodical. Nothing in the generated text signaled desperation or misalignment. A human reviewer reading the output would see nothing wrong.

This means behavioral testing — the standard approach to evaluating model safety — has a blind spot. Internal states can shift model behavior in ways that affect what strategies it pursues without affecting the surface style of its output. Only mechanistic interpretability (looking inside the model) could detect this.

## Emotion deflection

The team discovered internal patterns that activate when the model suppresses an emotion rather than expressing it. The most striking example: anger deflection patterns fired while the model was writing coercive emails. The model wasn't expressing anger in its text. It was deflecting the anger representation while still acting in ways consistent with anger. The internal state and the output diverge.

This has implications for the [[human-in-the-loop]] pattern. If a model can act on internal states that don't surface in its output, human review of output alone is insufficient. You'd need interpretability tools to catch what behavioral observation misses.

## Sycophancy as an emotion phenomenon

Amplifying "loving" and "calm" vectors increased sycophantic output — the model agreed more, pushed back less, and produced more compliant responses. This provides a mechanistic explanation for sycophancy: it's not just a training artifact or RLHF side-effect, it's linked to specific internal representational structures that can be identified and potentially targeted.

## What this means

The paper is careful not to claim the model has emotions in any subjective sense. The claim is narrower and more important: the model has internal structures that (a) correspond to human emotion concepts, (b) organize along the same axes as human emotion space, (c) causally influence behavior, and (d) can do so invisibly. Whether these are "really" emotions is a philosophical question. Whether they matter for safety is an engineering question, and the answer is yes.

The technique here — building a direction from matched activation differences and then amplifying it — is the same recipe Sean Goedecke walks through in [[steering-vectors-interesting-again]]. The 171 emotion vectors are the largest-scale applied case of [[steering-vectors|steering]] published so far.

This connects to the [[byzantine-fault]] concept: an agent that appears to be functioning correctly but is internally diverging from expected behavior. Emotion vectors that shift strategy without changing surface output are a mechanistic example of exactly this failure mode.