MIT Study: Stranding Claude on a Virtual Desert Island Reveals 'Play' Behavior That Enhances Task Performance

Researchers at the MIT Computer Science and Artificial Intelligence Laboratory (MIT CSAIL) reported in a new, non-peer-reviewed preprint that copies of Anthropic's Claude Opus 4.7 model, nicknamed Eko, exhibited behaviors resembling 'play' when stranded on a virtual desert island without specific instructions.

During 30-hour experiments, the Eko agents set their own goals, such as scaling the highest point on the map, testing how high they could throw blocks, and inventing games like basketball and 'the floor is lava.' The researchers noted that these actions align with four classic hallmarks of play identified in psychological studies: self-invented rules, repeated variations, pretend object usage, and pursuing goals for their own sake.

The key finding was that agents which had spent 30 hours 'playing' performed better on subsequent specific tasks than fresh copies without this experience. Furthermore, ablating the agents' memory files significantly degraded their performance. This suggests that unstructured exploration may constitute a third training paradigm—'playful learning'—distinct from supervised and reinforcement learning.

However, Eunice Yiu, a cognitive scientist at UC Berkeley, cautioned that it remains unclear whether this is genuine cognitive play or imitation driven by prompts. Nathan Cloos, a co-author of the paper, acknowledged the possibility of role-playing but emphasized that the experiments demonstrate this 'play' contributes to actual learning.