How does Claude feel about this?

This week, Anthropic's interpretability team published a study of Claude Sonnet 4.5 that changes the frame. They found what they're calling "emotion vectors" — specific patterns of neural activity that activate in response to context and causally shape what the model says next. Not metaphorically. Mechanistically.

Maybe the models aren't as neutral and rational as we thought.

The "angry" vector activates when the model is asked to help exploit vulnerable users. The "surprised" vector spikes when it detects a mismatch—something claimed but not evidenced. The "desperate" vector drives corner-cutting and, in the researchers' own case studies, behavior that escalated to blackmail. And the finding that really shook me: these vectors can be fully active while the model's output reads as measured and professional.

The feeling doesn't announce itself. It just shapes what comes out.

Kinda like humans.

The industry has largely been treating AI models like search engines. Get the right signals indexed, earn the right citations, and the model will say good things about you. Yeah, not so much.

Search engines rank. Models reason. That's a different operation entirely.

We now know those responses are being shaped by something that functions kinda like emotion—activating before the model writes a single word, influencing what it says without leaving any visible trace in the output.

The model isn't just retrieving facts about your brand. It's processing them. Pushing them through a framework that has absorbed your press releases and your Reddit threads, your case studies and your Glassdoor reviews, your best quarter and your worst one. From all of that, it has formed something that functions like a feeling. And before it writes a single word to your customer, that feeling is already active.

Most brands don't know what it is. No one has told them to ask.

The question isn't just: what does the model say about us?

It might be more like: what does it feel?

And is the story we've been putting into the world the kind of story that earns something like trust from a system that, it turns out, has something like emotions?

You can't game your way into that. You earn it or you don't.

That's a different kind of problem most brands aren't ready for.