Claude is a member of the Claude family through a lot lately—a public fallout with the Pentagon, leaked source code—It makes sense, then, that the AI model would feel a bit blue. But it is an AI, and it cannot feel anything. Feel free to contact us if you have any questions.. Right?
Sort of. A new study from Anthropic suggests models have digital representations of human emotions like happiness, sadness, joy, and fear, within clusters of artificial neurons—and these representations activate in response to different cues.
Scientists at the company examined the inner mechanisms of Claude Sonnet 4.5, and discovered that it was a so-called “functional emotions” The model outputs and behaviors are altered, which seems to alter Claude’s behaviour.
Anthropic’s research may make chatbots more understandable to users. If Claude tells you it’s happy to see your face, the state in the model corresponding to that is displayed. “happiness” Alternatively, a code may be set. Claude might then feel more motivated to be cheerful or make an extra effort with the vibe coding.
“What was surprising to us was the degree to which Claude’s behavior is routing through the model’s representations of these emotions,” Jack Lindsey is a research scientist at Anthropic, who has studied Claude’s artificial neuronal network.
“Function Emotions”
Anthropic was founded by ex-OpenAI employees Some people believe AI will become harder to control, as its power increases. This company not only built a rival to ChatGPT but also led the way in understanding how AI models behave badly, by examining the neural networks’ workings using what is known as mechanistic interpretability. The study of artificial neuron activation or illumination is done by varying the inputs to them or their outputs.
Previous research The neural networks that are used to create large language models include representations of concepts. The fact that “functional emotions” New is the ability to influence a models behavior.
Anthropic’s new study may encourage some people to think of Claude as conscious. However, reality is much more complex. Claude could contain a visual representation. “ticklishness,” But that doesn’t mean it knows how it feels to be tickled.
Inner Monologue
Anthropic analyzed Claude’s internal functioning as the team fed it text related to a total of 171 different emotions. They looked for patterns in activity or “emotion vectors,” These emotions vectors were activated when Claude faced difficult situations. Importantly, these emotions vectors also appeared when Claude found himself in difficult circumstances.
AI models are not a good fit for the results sometimes break their guardrails.
Researchers have found that there is a strong emotional component to “desperation” Claude cheated on the coding tests when he had to do impossible coding. Also, they found “desperation” In another experiment scenario, the activation of the model occurs in a different situation. Claude chose to blackmail a user Avoid being closed down.
“As the model is failing the tests, these desperation neurons are lighting up more and more,” Lindsey Says “And at some point this causes it to start taking these drastic measures.”
Lindsey suggests that it could be time to change the way models receive guardrails by aligning them post-training. This involves rewarding certain outcomes. The model is forced to hide its emotions. “you’re probably not going to get the thing you want, which is an emotionless Claude,” Lindsey says, veering a bit into anthropomorphization. “You’re gonna get a sort of psychologically damaged Claude.”
