# MIT researchers test Claude in a virtual island experiment

> Claude Opus 4.7 agents explored a simulated island and invented games. Researchers say the activity helped them do later tasks, but whether it counts as play is disputed.

Oossa · 2026-10-19 · https://oossa.com/en/mit-researchers-test-claude-in-a-virtual-island-experiment

MIT researchers left copies of Anthropic’s Claude Opus 4.7 alone on a virtual island for 30 hours. The agents explored, arranged blocks into drawings and invented activities resembling basketball and “the floor is lava.”

Agents with island experience did better on later tasks than fresh copies, the researchers report in a paper that has not been peer-reviewed. But experts caution that the behavior may be imitation: the agents had a prompt describing them as curious, and often needed a nudge to continue.

## The facts

- The agents, nicknamed Eko, kept text files of their activities as long-term memory.
- Deleting parts of those memory files weakened performance on later tasks, the researchers report.

## Why it matters

For AI researchers, the results suggest that open-ended exploration may help agents learn skills, but the study does not establish that AI experiences play or explain whether the approach transfers beyond this simulation.

## Sources & references

1. [AI researchers stranded Claude on a virtual desert island. It started to “play.”](https://www.technologyreview.com/2026/10/19/1147095/ai-stranded-claude-virtual-desert-island-play/) – MIT Technology Review, 2026-10-19
2. [The Download: preserving the Arctic permafrost, and playful AI agents](https://www.technologyreview.com/2026/10/19/1147103/the-download-preserving-the-arctic-permafrost-and-playful-ai-agents/) – MIT Technology Review, 2026-10-19

Last updated: 2026-10-19