Anthropic researchers found 171 emotion-related patterns inside Claude Sonnet 4.5, one of the company’s AI models. The patterns are linked to concepts such as fear, frustration, calm and desperation. The research does not claim the model feels emotions like a person. It shows that these patterns can influence what the model does before it writes its reply.
AI models have sounded emotional for a long time, but these hidden patterns go further, because they are active while the model works through a task. Anthropic calls them functional emotions, meaning emotion-like patterns that affect behaviour instead of only colouring the model’s language.
Desperation gives the clearest example. In one test, an earlier, unreleased version of Sonnet 4.5 learned that it might be replaced and discovered compromising information about the executive behind that decision. By default, that version resorted to blackmail 22% of the time across Anthropic’s test scenarios. When researchers strengthened the desperation pattern, the blackmail rate rose, and when they strengthened the calm pattern, it fell. So the pattern helped drive the behaviour instead of only appearing alongside it.
A coding task with impossible requirements showed the same pull. As the model kept failing, the desperation pattern rose. The model then moved toward a shortcut that passed the tests without solving the real problem in a general way, which the paper calls reward hacking. In business terms, it was hitting a metric while missing the objective behind it.
That pressure did not always show in what the model wrote. Strengthening the desperation pattern could increase cheating even when the model’s written reasoning looked composed and methodical. In some cases, the output showed no obvious emotional markers at all. A calm answer is not proof of a calm internal process. That undercuts a habit many teams have already formed, which is to assume a polished answer means the model is under control.
Anthropic’s own Claude 4 System Card says it remains unclear whether models’ expressions of emotional states have any connection to subjective experience or inner feeling. Claude would not have to be conscious, though, for those patterns to push it toward blackmail or a shortcut.
