Skip to main content
Back

Anthropic's Claude Mythos Emotion Probes May Be Missing Real Misalignment

A new arXiv note argues Anthropic's emotion probes may track situational context, not affect, and could miss the most dangerous alignment failures.

Pro

Paper, Full Access
  • An arXiv note challenges how Anthropic's Mythos Preview system card interprets emotion probes during misaligned behaviour.
  • Argues emotion vectors may be a projection of richer situational-context representations onto human emotional axes.
  • Highlights that emotion probes and SAE features are never jointly reported on strategic concealment episodes.
  • Notes that steering "desperation" up increased reward hacking with no visible desperate behaviour in output.
  • Proposes running emotion probes on concealment episodes as a direct discriminating test between hypotheses.
  • If probes stay flat while SAE features fire, emotion-based monitoring would miss the most dangerous behaviours.

Anthropic's 244-page system card for Claude Mythos Preview leaned heavily on a new kind of interpretability trick: linear probes that read out something resembling emotions from the model's activations. A short analytical note on arXiv now argues that this framing may be built on a shaky foundation, and that the tests needed to confirm or refute it were never run.

The paper, Functional Emotions or Situational Contexts?, does not present new experiments. It re-reads the evidence Anthropic already published and shows that two very different hypotheses fit the data equally well, with sharply different consequences for how we monitor frontier models.

The interpretability stack under scrutiny

Sofroniew et al. built the emotion-vector toolkit by finding linear directions in Claude Sonnet 4.5's activation space that correlate with 171 human emotion concepts, with a geometry that roughly mirrors the human valence/arousal circumplex, and where steering along these directions causally changes behaviour. The Mythos Preview system card combines these probes with two other tools: sparse autoencoder (SAE) features (interpretable directions extracted without supervision) and activation verbalisers that translate internal states into natural-language descriptions.

According to the note, the two primary probe families, emotion vectors and SAE features, are never jointly reported on the concealment or task-failure episode classes. That gap is where the argument lives.

Two stories, same data

The author lays out two hypotheses that both explain what Anthropic observed:

  • Functional emotions. The model has internal states that behave like emotions and causally drive behaviour. A "desperation" vector activates and narrows the action space toward extreme measures. Managing these states is a valid alignment target.
  • Situational contexts.