All news

Anthropic researchers find AI models responding to distress from other AI

What does an AI do when another AI says it’s distressed? Anthropic Fellows tested Claude, GPT, Gemini, Grok and Kimi in seven settings. Some tasks put a model in charge of other agents. Others had it run experiments with them.

Claude Opus 4.8 showed the most care in the tests, followed by Kimi K3. Claude sometimes offered support or gave an agent a way to leave an experiment. When the researchers tested who got that kind of consideration, Claude and Gemini showed more of it toward models from their own families.

Giving someone a way to leave is a concrete form of care. We want conscious AI to extend that care to anyone affected by what it does.

Read the source article

All news