Anthropic analyzed the internal mechanisms of Claude Sonnet 4.5 and found “emotion-related” representations that shape its behaviour in a fashion that echoes human psychology, with more similar emotions corresponding to more similar representations. In contexts where you might expect a certain emotion to arise for a human, the corresponding representations are active. For instance, patterns related to desperation can drive the model to take unethical actions such as blackmailing a human to avoid being shut down. Overall, it appears that the model uses functional emotions—patterns of expression and behavior modeled after human emotions, which are driven by underlying abstract representations of emotion concepts.
(Whether it is right to call these patterns “emotions” I will leave to philosophers, but note that the AI researchers are careful to say “emotion-like representations” in the paper, presumably meaning something like “internal states that play a functional role similar to emotions”. Your bot isn’t crying because it’s sad, it’\s just that certain patterns activate when it processes text with a sad context. Although I suppose from the functional perspective, whether the bot is showing “real” emotion may be practically irrelevant.)
Snippings 