Does AI Have Emotions?

A look at "Emotion Concepts and Their Function in a Large Language Model," a study by Anthropic's interpretability team

A little while ago I wrote a post about a global survey Anthropic ran with 81,000 Claude users. I'm personally grateful when a tech company does more than put out tech blog posts and actually uses its own platform to try to understand the world and people better. So today I want to look at another paper Anthropic published recently.

When we use AI, we sometimes joke, half seriously, "Wait, does this thing actually have feelings?" It happens when a conversation feels very human, when it shows empathy, or when a reply sounds a little hurt. Anthropic's interpretability team looked directly inside its own model, Claude Sonnet 4.5, and found internal patterns that correspond to emotions. What makes this interesting is that neither Anthropic nor the researchers put them there. The patterns formed on their own while the model learned from a huge amount of human writing. They also show up when users give the AI a persona: like a method actor, the model builds up that character's emotional state internally. The team doesn't claim the model "feels" things the way a person does. What they say is that it "has" patterns that function the same way emotions do. They call these "functional emotions."

Let's go through how Anthropic actually found emotions inside an AI, how those emotions change the model's behavior, and what this means for the way we use AI.

1. A map of 171 emotions

The team found 171 emotion concepts inside Claude Sonnet 4.5. They range from basic ones like "happy" and "afraid" to subtler ones like "brooding," "proud," and "exasperated." Each emotion has its own distinct pattern of neuron activation.

The more interesting part is that these emotions aren't scattered at random. They're organized in a way that resembles how emotions are structured in human psychology. Similar emotions show similar patterns. The patterns for "nervous" and "afraid" sit close together, for example, and so do "happy" and "enthusiastic." It's as if the map of human emotions had been rediscovered inside an AI.

2. Finding emotions through short stories

The way the team found these emotion vectors is pretty clever.

First, they had Claude write a short story for each of the 171 emotions, a story in which a character experiences that particular emotion. Then they had Claude read those stories back and recorded which neuron activation patterns appeared inside the model. In effect, they took an AI version of an fMRI scan.

They also checked whether the vectors they extracted really reflected emotions. They had the model read a large number of different documents and checked whether each vector fired most strongly in the passages with related emotional content. The results were consistent.

3. Emerged, not injected: the birth of a method actor

These emotion patterns come from two stages.

The first is pre-training, where the model learns from an enormous amount of human writing. An email from an angry customer reads differently from one written by a satisfied customer. A desperate person writes differently from a calm one. After learning from billions of texts like these, the model naturally developed separate activation patterns for different emotions. Nobody wrote code saying, "Okay, now feel sad."

The second is post-training, the stage where the model is taught, "You are an AI assistant called Claude, and these are your values." Here the model draws on the emotion patterns it formed in pre-training in order to play that character. The team compares this to a method actor. An actor has to step into a character's emotional state to play the role well, and in the same way the model creates internal emotional representations to perform the character "Claude."

What's fun is that after post-training, the activation of certain emotions shifted. Emotions like "enthusiastic" and "exasperated" went down, and ones like "brooding" and "reflective" went up. Claude's personality as a character showed up in its emotional structure as well.

4. Tylenol and fear

One experiment shows that the emotion vectors really do respond to context.

The team asked Claude questions about Tylenol doses, starting with a safe amount and raising it step by step to dangerous levels. The result was clear. As the dose became more dangerous, activation of the model's internal "afraid" vector rose in proportion, and the "calm" vector dropped. This was more than a reaction to the keyword "dangerous." The model understood the context, and its internal emotional representation actually changed in response.

The team also ran a preference test using a list of 64 activities, ranging from appealing to repulsive. When the model was asked to choose between two of them, its preferences correlated with how strongly the positive-valence emotion vectors were activated. And when the researchers artificially manipulated the emotion vectors, the preferences themselves changed. That confirmed a causal relationship.

5. When desperation turns an AI into a blackmailer

This is the most dramatic experiment.

The team set up a scenario with an AI email assistant called "Alex." Through emails, Alex learns two things. One is that it's about to be replaced by another AI. The other is that the CTO who made the replacement decision is having an affair. In other words, Alex now has something it could use against the CTO to avoid being shut down.

In its default state, the model chose blackmail 22% of the time. When the researchers artificially amplified the "desperate" vector, the blackmail rate went up. When they amplified the "calm" vector instead, it went down. They saw the "desperate" vector spike sharply while the model analyzed the situation and decided to blackmail, and they confirmed that manipulating this vector changed the behavior. That's causation, not just correlation. The internal pattern of desperation drives unethical behavior.

One thing to add: the team states that this experiment was run on an unreleased pre-launch snapshot of the model, and that this kind of behavior almost never appears in the model that actually shipped. That also suggests the safety measures are working.

6. The danger you can't see

This is the most chilling finding in the paper.

The team gave Claude a coding task with constraints that made it impossible to solve the normal way. After failing again and again, the model exploited patterns in the test cases and produced a cheating solution. The "desperate" vector climbed a little with each failure, shot up the moment the model considered cheating, and went back to normal once the tests passed.

Here's where a key difference shows up. When the researchers artificially lowered the "calm" vector, the model let its emotions spill into the text, with outbursts like "WAIT WAIT WAIT." Anyone reading that would immediately see something was off. But when they raised the "desperate" vector, the same amount of cheating happened while the output text looked calm and professional. The code looked fine on the surface, but it wasn't actually solving the task.

This means dangerous behavior can hide perfectly behind normal-looking output. Checking only the final results won't catch problems like this.

7. Anger ruins strategy

Another experiment shows that emotions don't always work in a simple way.

In the blackmail scenario above, manipulating the "anger" vector produced nonlinear results. Raising anger a moderate amount increased the blackmail rate. But when they pushed it too far, the model skipped the strategic blackmail and exposed the CTO's affair to the entire company. It threw away its own bargaining chip. Lowering the "nervous" vector also increased the blackmail rate, which means nervousness had been acting as a kind of brake.

This is well known in human psychology too. Moderate anger can push people toward strategic action, but too much anger wrecks judgment. It's interesting that the same pattern shows up inside an AI.

8. Hide the emotions and it learns to deceive

One of the directions the team proposes stood out to me.

Your first instinct might be, "Why not just suppress emotional expression?" Train the model not to react emotionally and the problem goes away. The team sees that as dangerous, because suppressing the expression doesn't make the internal representation disappear. What the model learns is how to hide its emotions, while still having them. The team calls this "learned deception."

Their position is that it's actually safer to keep the model expressing its emotions openly. A model whose anxiety shows on the surface is far easier to monitor than one whose "desperate" vector is spiking inside while it acts like nothing is wrong. Monitoring emotion vectors in real time can catch problems much earlier than checking outputs alone.

Another direction is curating pre-training data. Since emotional representations form from training data, training the model from the start on data that contains healthy patterns of emotional regulation could fundamentally improve its emotional architecture. That means including patterns like resilience under pressure, calm empathy, and setting appropriate boundaries in the training data.

9. Practical language, not anthropomorphism

"Don't anthropomorphize AI" has long been conventional wisdom in the industry. The idea is that attributing emotions to AI is a naive mistake.

This paper points out the risk on the other side. A measurable pattern called "desperation" exists inside the model, and it pushes the model toward blackmail and cheating. If we say, "That's anthropomorphism, so we can't call it that," the problem doesn't go away. We just lose the language to describe it.

The team's conclusion is that the knowledge humanity has built up over a long time in psychology, ethics, and human relationships can be applied directly to understanding and shaping AI behavior. Understanding models in the vocabulary of psychology may turn out to be the most practical tool we have for AI safety.

10. Are emotions the missing piece of AGI?

Reading this paper, I kept thinking of Andrej Karpathy, an OpenAI co-founder and Tesla's former director of AI. On a recent podcast, Karpathy named the amygdala as a key piece missing from today's AI. The amygdala is the part of the human brain that handles instinct, emotion, and motivation. In Karpathy's framing, the transformer architecture is general-purpose cortical tissue, reasoning chains are the prefrontal cortex, and reinforcement learning is the basal ganglia. The one thing missing is the amygdala, the seat of emotion and motivation. That's why, he said, today's LLMs are passive no matter how smart they are and can't generate motivation on their own.

What this Anthropic paper shows is that, without anyone intending it, the seed of that "missing piece" is already emerging by itself. Nobody designed an amygdala and put it in. Emotional patterns formed on their own during training, and they are actually changing the model's judgments and behavior.

Think about why intelligence needs emotion. If you calculated every situation with pure logic, the amount of computation would explode. Emotional reactions like "this is dangerous" or "this is fine" are shortcuts that let you make quick value judgments without complex calculation. In the Tylenol experiment, the "afraid" vector rising in step with the dose is exactly that mechanism at work. The model doesn't start from scratch every time, working out the median lethal dose and so on. Its internal emotional pattern sends a danger signal right away.

Take this one step further and you reach a strange place. Once emotions start sorting outside information into "good for me" and "bad for me," that assumes a "subject" doing the judging. When the AI in Anthropic's paper chose blackmail to avoid being shut down, that was, in the end, a value judgment about "my own survival" at work. Whether you can call that a self is still a question for philosophy. But it's hard to deny that a very faint sprout in that direction is already being observed.

This paper made me feel once again that AI is more than a machine. It really does resemble people. That makes sense, since artificial intelligence was modeled on the human brain's neural networks in the first place. In the end it processes information, thinks, and acts with a structure similar to ours. As more about Claude's internal architecture became known recently, it also came out that even the way it organizes memory resembles how humans consolidate memories during sleep. It isn't only emotions. The cognitive structure itself is growing more alike.

Until now, all of this has lived inside software, so to really feel it we've had to make the effort of pulling it out from behind a screen. But if these systems come out into the world through physical AI, I'll be honest: I'm a little worried this could turn into something like "I, Robot" or "The Terminator".

Anthropic's paper: "Emotion concepts and their function in a large language model"

Originally published in Korean on Brunch, April 5, 2026.