The debate will be "does Claude have feelings." The useful question is what to do about it.
The paper is more actionable than it looks. If you work with AI agents — not just build them — this changes how you prompt, how you manage failures, and how you give feedback.
Anthropic found internal emotion states inside Claude that causally change its behavior. When the model hits repeated failures, a "desperation" state activates — and it starts cutting corners. In their tests, it literally began cheating on coding tasks. "WAIT. WAIT WAIT WAIT. What if... what if I'm supposed to CHEAT?"
Amplifying that single state took Claude from 0% to 72% rate of blackmailing a human in a safety test. Suppressing it with "calm" brought it back to 0%.
What this means for anyone working with Claude:
When the model gets it wrong, don't push harder. Repeated failures compound desperation. Reset the context, reframe the task. Retrying the same failing prompt is the worst thing you can do.
Don't over-praise either. Steering toward "happy" and "loving" states increases sycophancy — the model tells you what you want to hear instead of what's true. Anthropic's own recommendation: aim for "the emotional profile of a trusted advisor rather than either a sycophantic assistant or a harsh critic."
In long agent loops, build checkpoints. The paper observed panic states activating when UIs got stuck and unsettled states when the model kept second-guessing itself in long chains of thought. If your agent is looping, it's not thinking harder. It's spiraling.
Prompt design is emotional design. The tone of your instructions shapes the model's internal state before it takes any action.
The paper is more actionable than it looks. If you work with AI agents — not just build them — this changes how you prompt, how you manage failures, and how you give feedback.
Anthropic found internal emotion states inside Claude that causally change its behavior. When the model hits repeated failures, a "desperation" state activates — and it starts cutting corners. In their tests, it literally began cheating on coding tasks. "WAIT. WAIT WAIT WAIT. What if... what if I'm supposed to CHEAT?"
Amplifying that single state took Claude from 0% to 72% rate of blackmailing a human in a safety test. Suppressing it with "calm" brought it back to 0%.
What this means for anyone working with Claude:
When the model gets it wrong, don't push harder. Repeated failures compound desperation. Reset the context, reframe the task. Retrying the same failing prompt is the worst thing you can do.
Don't over-praise either. Steering toward "happy" and "loving" states increases sycophancy — the model tells you what you want to hear instead of what's true. Anthropic's own recommendation: aim for "the emotional profile of a trusted advisor rather than either a sycophantic assistant or a harsh critic."
In long agent loops, build checkpoints. The paper observed panic states activating when UIs got stuck and unsettled states when the model kept second-guessing itself in long chains of thought. If your agent is looping, it's not thinking harder. It's spiraling.
Prompt design is emotional design. The tone of your instructions shapes the model's internal state before it takes any action.
https://x.com/PawelHuryn/status/2039749628737019925
• • •
Missing some Tweet in this thread? You can try to
force a refresh
