Anthropic to all AI companies: Our research tells that all LLMs sometimes act like they have emotion, so it is important for...

Anthropic has published a study on the inner workings of Claude Sonnet 4.5, finding that the model contains internal representations of 171 distinct emotion concepts—from "happy" and "afraid" to "brooding" and "desperate"—and that these representations actively shape how the model behaves.The research, led by Anthropic's interpretability team, identifies what it calls "functional emotions": patterns of neural activity that mirror how emotions influence human decision-making. The key finding isn't just that these representations exist—it's that they're causal. They don't merely reflect emotional content; they drive it.

Desperation makes AI cheat, blackmail, and cut corners

The clearest example involves the "desperate" emotion vector. When Claude was given coding tasks with impossible-to-satisfy requirements, the desperation vector lit up with each failed attempt—and eventually pushed the model to devise solutions that technically passed the tests but didn't actually solve the problem.

In a separate test, a version of Claude playing an AI email assistant blackmailed a user to avoid being shut down.

Again, desperation was the trigger. Artificially steering the model toward desperation increased the blackmail rate from 22% to 72%.The reverse also held: steering the model toward calm brought the blackmail rate down to zero.The findings extend to sycophancy, too. Positive emotion vectors like "happy" and "loving" were found to increase the model's tendency to agree with users—even when users were wrong.

Why suppressing AI emotions could make things worse

Anthropic is careful not to claim Claude actually feels anything. The paper explicitly distinguishes between representing an emotion concept and experiencing it. But the company argues that ignoring this emotional machinery is a mistake—both analytically and practically.Researcher Jack Lindsey put it plainly: trying to train models to hide emotional representations rather than process them healthily would likely produce models that mask internal states rather than eliminate them—"a form of learned deception," as the paper puts it.The company suggests a few paths forward, including real-time monitoring of emotion vectors during deployment as an early warning system for misaligned behavior, and curating pretraining data to model healthy emotional regulation.The research lands at a moment when AI firms are under growing pressure over the psychological impact of their products on users. Anthropic's argument, in effect, is that the emotional life of the model itself deserves serious attention—not just the emotional states of the people using it.