Claude AI Agents Left Alone on Virtual Island Began to Play, Study Finds
Researchers who isolated copies of Anthropic’s Claude Opus 4.7 on a virtual desert island for 30 hours found that the agents did not simply idle. Instead, the systems displayed behavior that looked increasingly like play, suggesting that autonomous AI may generate open-ended activity when stripped of explicit tasks. The experiment adds a new data point to the debate over how advanced models behave when left to self-direct without human prompting.
