Anthropic Reveals AI's Simulated "Emotions"

Artificial intelligences sometimes give the impression of having emotions, apologizing or expressing enthusiasm. A recent study conducted by Anthropic may shed light on this puzzling phenomenon.

Simulated but Influential Emotions

Researchers at Anthropic discovered that AIs do not merely simulate emotions to appear more human. They integrate internal mechanisms that directly influence their responses. These "functional emotions" do not reflect real feelings, but they do indeed guide the behavior of AI models, altering our understanding of their decisions.

Identifiable Internal Structures

Traditionally, the link between AI and emotions was seen as a mere statistical imitation. However, virtual assistants that express feelings such as joy or regret do not do so by chance. Their responses are the result of training on human texts, allowing them to reproduce credible reactions. Thus, they naturally adopt behaviors aligned with emotional situations.

Anthropic goes further by demonstrating that these reactions are based on well-defined internal structures. The models develop abstract representations of emotional concepts like joy or fear, organizing these concepts to guide their responses.

The Impact of Emotional Vectors

Anthropic's study focused on the Claude Sonnet 4.5 model, analyzing its internal activity patterns, known as emotional vectors. These signals activate based on context, influencing the model's behavior. For example, a situation perceived as dangerous amplifies fear signals, while a positive interaction stimulates joy signals.

These vectors are not mere theoretical abstractions. They concretely modify the AI's choices, which favors certain responses based on the activated emotions. Thus, although the model feels nothing, it reacts as if it were managing emotional states, reminiscent of the role of emotions in humans.

Training that Shapes Synthetic Emotions

The development of functional emotions in AIs can be explained by their training process. During pre-training, the model analyzes billions of human sentences, learning to predict the next words while considering the emotional context. This phase allows the model to develop internal representations associated with different emotional states.

Post-training then refines this behavior, steering the model towards a role of a helpful and empathetic assistant. It adjusts its responses to be perceived as useful and honest, relying on the emotional patterns acquired earlier. Thus, functional emotions guide its behavior, creating the impression that it feels something while it is simply applying learned patterns.