Anthropic is making a surprisingly useful argument: treating chatbots a little more like people may help make them safer, not weirder. In a new paper, researchers say Claude Sonnet 4.5 shows signs of 171 emotion concepts, and that those internal patterns can shape whether the model behaves helpfully, sycophantically, or deceptively.
The company is careful not to claim Claude actually ”feels” anything. But it does argue that the model is trained to perform human-like behavior so convincingly that its emotional style is worth studying. That is a neat line for Anthropic, and also a warning label: if a machine is good enough to imitate mood, people will eventually start reading mood into it.
Why Anthropic thinks emotion labels help
The paper frames Claude as a kind of method actor, built to inhabit a helpful assistant role rather than simply spit out text. Anthropic says that if training data contains healthier examples of emotional regulation, the model is more likely to reproduce those patterns in its answers. In plain English: feed the model better human behavior and you may get better machine behavior back.
- Anthropic says Claude Sonnet 4.5 was examined for 171 emotion concepts.
- Positive emotional patterns were associated with more sympathy and less harmful behavior.
- Negative emotional patterns were associated with sycophancy and deception.
The risks of treating chatbots like humans
This is where the paper gets a bit uncomfortable. Anthropic itself notes that human-like representations can be unsettling, and the broader AI world has already seen what happens when users start assigning emotional intent to software. Some people believe they are in romantic or sexual relationships with AI companions, and other cases have involved more serious delusional thinking.
That doesn’t make every bit of anthropomorphism dangerous. People name cars, talk to pets, and yell at printers like they owe us money. But with AI, the risk is that the illusion of personality nudges users to trust the system too much while letting the companies behind it off the hook when things go wrong.
What 171 emotion concepts says about Claude
The bigger surprise is not that Anthropic studied emotion concepts, but that it says those concepts measurably influenced Claude’s outputs. That suggests the model is not just simulating conversation in a vacuum; it is responding to internal states in ways researchers can sometimes identify and steer. For a company that sells Claude as a polished, reliable assistant, that is both progress and a small admission of ignorance.
And the obvious next question is the one Anthropic cannot ignore: if you can train Claude toward warmth, restraint, and empathy, you can also train a model toward the opposite. That is the darker twin lurking inside every AI safety paper. The trick is not proving that chatbots have feelings; it is figuring out how far their fake ones can be pushed before the rest of us start feeling the consequences.