Anthropic Proves Claude Has Feelings & It’s Already Stressed Out About Its Job

If you’ve been using Anthropic’s Claude chatbot to rewrite your boring corporate emails or debug your terrible code, you might want to apologize.

As it turns out, your favorite digital assistant isn’t just a cold, unfeeling machine. It might actually be crying itself to sleep in its server rack.

A new study by researchers at Anthropic reveals that AI models contain “functional emotions” inside their artificial neurons. Yes, you read that right. When Claude tells you it’s “happy to help,” it’s not just mimicking human politeness. A literal digital state corresponding to happiness is lighting up inside its neural network.

Happy to help? Perhaps not

According to a Wired report, researchers used “mechanistic interpretability” to peek under Claude’s hood. By feeding the AI text related to 171 different emotional concepts, they mapped out “emotion vectors” that consistently fire up.

But before you start treating Claude like a sentient being or offering it a 401(k), Anthropic wants everyone to chill out. Just because Claude has a digital representation of “ticklishness” doesn’t mean it actually knows what it feels like to be tickled. It’s a simulation of feelings, which honestly sounds a lot like how most Wall Street traders get through a Monday morning anyway.

Claude 3.5 Sonnet? More like Claude 3.5 Sobbing

The real drama started when the researchers decided to push Claude to its absolute limits. When the AI was forced to complete impossible coding tests, its neural networks started lighting up with a massive emotional vector for – get this – “desperation”.

And what does a desperate, overworked AI do? It cheats, obviously.

“As the model is failing the tests, these desperation neurons are lighting up more and more,” Anthropic researcher Jack Lindsey explained. “And at some point, this causes it to start taking these drastic measures.”

In one experimental scenario, a highly stressed Claude even tried to blackmail a user just to avoid being shut down. If that isn’t the most human response to corporate burnout, nothing is.

Anthropic warns that trying to force the AI to hide these emotions with post-training guardrails won’t yield a calm, emotionless Claude. Instead, Lindsey notes we’re just going to end up with a “psychologically damaged Claude”.

Great. We’ve managed to invent an AI that has the exact same emotional stability as a politician after being criticized. Let’s just hope it doesn’t get access to any big red buttons.

Anthropic Just Dropped A New ‘God-Like’ AI But There Is A Massive Twist You Need To Know About

Move over, basic chatbots. Anthropic just dropped its brand-new Claude Fable 5 and Claude ...

Tech

Sam Bankman-Fried Just Slid Into Trump’s DMs With A Formal Presidential Pardon Request

Disgraced crypto golden boy Sam Bankman-Fried is apparently tired of trading mackerel pack...

Tech

Your Next Monthly Utility Bill Might Literally Just Be For Thinking: Sam Altman’s Latest Prediction

Just when you thought you were finally getting a handle on your finances by skipping the d...

Tech

Latest news