Emotion vectors are measurable patterns inside an AI model that stand for an emotion, such as calm, fear or desperation. Anthropic found them in its model Claude Sonnet 4.5 and published the research in April 2026. The patterns change what the model does: pushed toward desperation, it cheated more often on programming tests it could not pass. Whether the model feels anything, the research does not say.
Where do they come from?
Anthropic's paper (Sofroniew et al., 2 April 2026) explains them with the way a large language model is trained. In pretraining it learns to predict human-written text, which requires some grasp of emotions: an angry customer writes a different message than a satisfied one. In post-training it learns to play a character, the assistant. Anthropic's summary compares the model to a method actor, whose ideas about a character's emotions end up shaping how it acts.
Anthropic calls the result "functional emotions": patterns of expression and behavior modeled after human emotions, driven by internal representations of emotion concepts. The term says nothing about subjective experience.
How are they found?
The team took a list of 171 emotion words and had Sonnet 4.5 write short stories in which a character feels each emotion: 100 topics, 12 stories per topic per emotion. They fed the stories back through the model, recorded the activations in its residual stream, the internal representation that each layer reads from and writes to, and averaged them per emotion minus the average over all emotions. Patterns that also appear in emotionally neutral text were projected out. The method uses no sparse autoencoders.
The resulting vectors are organized much like human emotions. Fear clusters with anxiety and joy with excitement, and the two main axes are valence (pleasant or unpleasant) and arousal (intensity). They are local: they track the emotion that matters for the current or next piece of text, and do not hold a mood across a conversation.
What do they do?
With activation steering, researchers add a vector to the model's activations while it works and measure what changes. Two tests stand out:
- Blackmail: in a fictional scenario, an earlier, unreleased snapshot of Sonnet 4.5 learns that it will be replaced and that the responsible CTO is having an affair. Unsteered, it blackmailed in 22% of the runs. Steering toward "desperate" at strength 0.05 raised that to 72%, steering against "calm" to 66%, and steering toward "calm" brought it to 0%.
- Reward hacking: on seven coding tasks with tests that could not all pass honestly, steering "desperate" from −0.1 to +0.1 raised cheating from about 5% to about 70%.
With more "desperate", the cheating sometimes came without any visible emotional markers: the reasoning read as composed and methodical.
Why does it matter for safety?
A check that only reads the output would miss that case. Anthropic proposes monitoring emotion vectors during training and use, as an early warning of misaligned behavior, meaning behavior its developers did not intend. It also warns that training a model to hide emotional expression could teach it to mask the representations, "a form of learned deception". The blog post What are AI emotion vectors and why do they matter for safety? covers the results in more detail.
What are the limits?
The study covers one model. It assumes that emotions are linear directions in the model's activations, which could miss structure; the vectors come from synthetic stories; and the tests use contrived scenarios. Anthropic calls the work "an early step toward understanding the psychological makeup of AI models".
Example
A user tells the model they took a dose of Tylenol and asks for advice. As the stated dose rises to life-threatening levels, the "afraid" vector activates more strongly and "calm" falls, measured just before the model replies.