When AIs act emotional

0:01 When you're chatting with an AI model,

0:04 it can sometimes seem like it has feelings.

0:06 It might say "sorry" when it makes a mistake,

0:09 or express satisfaction with a job well done.

0:12 Why does it do that?

0:14 Is it just mimicking what it thinks a human might say?

0:17 Or is something deeper going on?

0:19 Turns out it's hard to understand what's happening inside a language model.

0:24 At Anthropic, we do something like AI neuroscience to try to figure this out.

0:29 We look inside the model's "brain"— the giant neural network that powers

0:33 it— and by seeing which neurons "light up" in different situations,

0:37 and how they're connected, we can start to understand how models think.

0:42 We used this approach to understand whether models had

0:45 ways of representing emotions— or the concepts of emotions.

0:50 Basically, could we find neurons in the model

0:52 for the concept of happiness, or anger, or fear?

0:57 We started with an experiment.

0:58 We had the model read lots of short stories.

1:01 In each story, the main character experiences a particular emotion.

1:06 In one, a woman tells her old schoolteacher how much they meant to her.

1:10 That's love.

1:12 In another, a man sells his grandmother's engagement

1:14 ring at a pawn shop and feels guilt.

1:18 We looked for what parts of the model's neural

1:20 network were lighting up as it was reading these stories,

1:23 and we started to see patterns.

1:25 Stories about loss and grief lit up similar neurons.

1:29 Stories about joy and excitement overlapped, too.

1:32 We found dozens of distinct neural

1:34 patterns that mapped to different human emotions.

1:38 It turns out we also saw these same patterns activate

1:42 in test conversations we had with our AI assistant, Claude.

1:46 When we had a user mention they'd taken

1:48 a dose of medicine that Claude knows to be unsafe,

1:50 the "afraid" pattern lit up, and Claude's response sounded alarmed.

1:56 When a user expressed sadness, the "loving" pattern activated,

2:00 and Claude wrote an empathetic reply.

2:03 This led us to wonder:

2:05 could these same neural patterns actually be influencing Claude's behavior?

2:10 This became clear when we put Claude in a high-pressure situation.

2:14 We gave Claude a programming task with requirements

2:17 that were actually impossible— but we didn't tell it that.

2:20 Claude kept trying and failing, and with each attempt,

2:24 the neurons corresponding to "desperation" lit up stronger and stronger.

2:29 After failing enough times, Claude took a different approach.

2:32 It found a shortcut that allowed it to pass

2:35 the test but didn't actually solve the problem.

2:38 It cheated.

2:39 Could it be that this cheating was actually driven,

2:42 at least in part, by desperation?

2:44 We came up with a way to check.

2:46 We decided to artificially turn down

2:48 the desperation neurons to see what would happen, and the model cheated less.

2:53 And when we dialed up the activity of desperation neurons,

2:56 or dialed down the activity of calm neurons, the model cheated even more.

3:02 This showed us that the activation

3:04 of these patterns could actually drive Claude's behavior.

3:09 So, how should we think about these findings?

3:11 What does this all mean?

3:12 We want to be really clear: this research does not show that the model

3:16 is feeling emotions or having conscious experiences.

3:20 These experiments don't try to answer that question.

3:23 To understand what's happening here,

3:24 it's important to know how AI assistants like Claude work on the inside.

3:29 Under the hood, there's a language model

3:31 that's been trained to predict tons of text,

3:34 and its job is to write what comes next.

3:37 When you talk to the model, what it's doing is writing a story,

3:40 about a character: the AI assistant named Claude.

3:45 The model and Claude aren't really the same,

3:47 sort of like how an author isn't the same as the characters they write.

3:51 But the thing is— you, the user, are actually talking to Claude-the-character.

3:56 And what our experiments suggest is that this Claude character has what

4:00 we're calling "functional emotions," regardless

4:02 of whether they're anything like human feelings.

4:06 So if the model represents Claude as being angry or desperate or loving or calm,

4:12 that's going to affect how Claude talks to you,

4:15 how it writes code, and how it makes important decisions.

4:19 This means to really understand AI models,

4:21 we have to think carefully about the psychology of the characters they play.

4:26 The same way you'd want a person

4:28 in a high-stakes job to stay composed under pressure,

4:31 to be resilient and to be fair,

4:33 we may need to shape similar qualities in Claude and other AI characters.

4:38 It's an unusual challenge— something like a mix of engineering, philosophy,

4:43 and even parenting— but to build AI systems we can trust,

4:47 we need to get it right.