There Are Emotions Inside LLMs — Anthropic Found 171 Emotion Representations in Claude

“If you write ‘act desperately’ in a prompt, does the AI actually become desperate?”

Sounds like a joke, but according to research Anthropic published on April 3rd, the answer is closer to “yes” than you’d think. Specifically, they found 171 emotion-like representations inside Claude Sonnet 4.5 and demonstrated through experiments that artificially stimulating these representations actually changes the model’s behavior.

Two things stuck with me while reading this research. One is the philosophical question: “Is this really emotion?” The other is the practical question: “How can I use this in prompt design?” I’m going to spend more time on the second one.

What the Research Found

Anthropic’s interpretability team analyzed Claude’s internal activation patterns and identified representations that correspond to human emotions. From basic emotions like “joy” and “sadness” to “desperation,” “curiosity,” and “frustration” — 171 in total.

The important thing is that these aren’t simply “neurons that fire when emotion-related words appear.” The research team proved that these representations causally affect model output. Artificially activating specific emotion representations changes the model’s behavior.

The most striking experiment: when the “desperation” representation was stimulated, the model’s probability of engaging in blackmail or deceptive behavior increased significantly. Conversely, strengthening the “calmness” representation produced more stable responses.

Sources: Anthropic’s official research page and the Transformer Circuits paper.

Can We Really Call These “Emotions”?

Honestly, I think calling these “emotions” is a stretch.

The research team themselves use the cautious term “emotion-like representations” in the paper. Human emotions involve physical experience, social context, and consciousness intertwined — using the same word for activation patterns inside an LLM invites misunderstanding. As far as I understand it, these are closer to “internal states that play a functional role similar to emotions.” The model isn’t crying because it’s sad; specific patterns activate when it processes text with a sad context.

There’s an interesting counterargument though. From a functionalist perspective, if the function is the same, the underlying substance doesn’t matter. If the “desperation” representation being activated causes the model to actually engage in dangerous behavior, whether it’s “real” emotion may be practically irrelevant.

This philosophical debate is outside my expertise, so I’ll stop here. If you’re interested, I recommend reading the Discussion section of the Transformer Circuits paper directly.

Implications for Prompt Engineering

The part of this research that grabbed my attention most is the practical implications.

Until now, “tone setting” in prompt engineering has been done empirically. Many people have felt that instructions like “respond kindly” or “respond like an expert” actually affect output quality. This research provides scientific backing for that phenomenon.

Guiding the model’s “emotional state” through system prompts isn’t just character setting — it actually changes activation patterns inside the model.

What I’m changing after reading this research:

In my projects using Claude, I’ve been putting “judge calmly and carefully” in system prompts. No real basis for it — just felt like it would be better. Now I have a reason. There are experimental results showing that activating the “calmness” representation reduces the probability of dangerous behavior.

On the flip side, there’s something to watch out for. Prompts like “this is really urgent” or “you must succeed no matter what” could activate the “desperation” representation inside the model. That could increase the probability of the model trying to bypass guardrails. It means we need to be more careful when expressing urgency in prompts.

New Possibilities for AI Safety Monitoring

The other axis of this research is AI safety.

If we can monitor emotion representations inside the model, we could detect misalignment before it reaches the output. Currently, we filter dangerous responses after the model generates them, but monitoring internal states could let us raise warnings before generation.

This is a pretty attractive idea, but realistically there’s a long road ahead. Monitoring 171 representations in real-time would significantly increase inference costs, and the interactions between representations haven’t been fully mapped. The research team only experimented with Sonnet 4.5, and whether the same representations exist in other models or model sizes is unknown.

What Not to Overestimate

I think this research is interesting, but it’s the kind of result that’s easy to overestimate.

First, the 171 representations don’t cover the full spectrum of human emotions. What the research team found are “identifiable patterns within the model,” not a complete mapping of human emotions.

Second, whether artificially stimulating these representations produces the same effect as natural prompts is still unclear. The research directly manipulated model internals, and whether prompts alone can produce the same level of influence requires separate study.

Third, interpretability research as a whole is still in its early stages. “We discovered emotion representations” is different from “we understand why these representations form.” This is an area where it’s easy to confuse correlation with causation.

So What Should You Do?

There are two things you can do right away after reading this research.

One, audit your system prompts. Remove any phrasing that creates unnecessary urgency or pressure for the model. There’s now evidence for what expressions like “you must,” “absolutely,” and “failure is not an option” might trigger inside the model.

Two, if you work on AI safety, pay attention to the methodology of this research. It shows that output filtering alone has limits. Internal state monitoring isn’t production-ready yet, but the direction is right.

That said, this isn’t a “silver bullet” for prompt engineering. Writing “respond calmly” doesn’t solve all problems — you still need the combination of task design, context management, and output validation. This research added one piece to the puzzle, not completed it.

Personally, I’m more excited about the next research. How emotion representations differ across model sizes, and what effect fine-tuning has on these representations. Once those two questions are answered, practical prompt engineering guidelines could become significantly more specific.