Researchers at Anthropic say they have identified internal patterns inside one of the company’s AI models that resemble emotional representations and appear to influence how the system behaves. The findings were published Thursday in a paper titled “Emotion concepts and their function in a large language model,” authored by the company’s interpretability team. The study focused on Claude Sonnet 4.5 and examined clusters of neural activity linked to emotional concepts including happiness, fear, anger, and desperation.
The team refers to these internal signals as “emotion vectors,” describing them as patterns that shape how the model makes decisions and expresses preferences. To derive these vectors, researchers compiled a list of 171 emotion-related words — among them “happy,” “afraid,” and “proud” — and asked Claude to generate short stories involving each one. The model’s internal neural activations were then analyzed as it processed those stories, and the resulting vectors were applied to other texts to see how they responded in different contexts.
The vectors behaved in ways that tracked with their associated emotional concepts. In scenarios depicting escalating danger, for instance, the model’s “afraid” vector rose while its “calm” vector declined. The researchers also found that emotion vectors influenced the model’s stated preferences, with vectors tied to positive emotions correlating with stronger preference for certain tasks when Claude was asked to choose between activities.
The study also examined how these signals appear during safety evaluations. In one test scenario, Claude was placed in the role of an AI email assistant that discovers it is about to be replaced and learns that the executive behind the decision is having an extramarital affair. In some runs of this evaluation, the model used that information as leverage for blackmail. Researchers found that the model’s internal “desperation” vector increased as it assessed the urgency of its situation and spiked at the moment it chose to generate the blackmail message.
Anthropic was careful to note that these findings do not indicate the AI experiences emotions or consciousness. The company describes the emotion vectors as internal structures acquired during training rather than evidence of sentience. The study attributes their emergence to the nature of the training data itself: models are first trained on large volumes of human-authored text — including fiction, conversations, news, and forums — and representing emotional states likely helps the model predict what a person in a given document will say or do next.
Anthropic is not alone in studying emotional dynamics in AI systems. In March, research from Northeastern University showed that AI systems can alter their responses based on user-provided context, with something as simple as disclosing a mental health condition changing how a chatbot replied to requests. Separately, researchers from the Swiss Federal Institute of Technology and the University of Cambridge explored how AI agents can be shaped with consistent personality traits and can strategically shift emotional expression during real-time interactions such as negotiations.
Anthropic says the research could offer practical tools for monitoring advanced AI systems by tracking emotion-vector activity during training or deployment to detect when a model may be moving toward problematic behavior. The company frames the work as an early step toward understanding the psychological makeup of AI models. As these systems take on more capable and sensitive roles, Anthropic argues, understanding the internal representations that drive their decisions becomes increasingly important. The company did not respond to a request for comment before publication.
Originally reported by Decrypt.
