MIT Stranded an AI on a Virtual Island for 30 Hours. It Played.

Thirteen AI agents spent 30 hours each alone on a virtual desert island, and instead of idling, they started to play — a result MIT researchers say could reshape how machines are trained.

The experiment comes from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) in a study released earlier this month, titled "Is this machine playing?" The setup was deliberately minimal. The researchers placed an agent named Eko, powered by Anthropic's Claude, on a bounded virtual island containing almost nothing: a large rock, some small blocks, and a golden ring, plus minor geographic features and water on all sides. Every item was meant to serve as a potential tool or toy.

The point was to remove instructions entirely.

What did the agents actually do?

The Ekos moved fast. They first learned to control their virtual bodies — stylized as bipedal lobsters, an apparent reference to the AI program OpenClaw — and then began inventing activities and goals to fill their time. The researchers ran 13 iterations of Eko for 30 hours each, observing behavior and reading the agents' trains of thought.

The reported antics read like a field notebook:

• One Eko stacked blocks into a pile and climbed to the island's highest point.
• Another lost most of its interactable items in the water — then held a quasi-memorial for its lost tools.
• A third built a Stonehenge-like monument; when it collapsed, the agent named the wreckage "Hubris."

Nathan Cloos, a PhD candidate at MIT and coauthor of the study, told Gizmodo the design philosophy behind the experiment: "We had this mindset to have as [few] constraints as possible, because we really wanted to see what the behavior of these autonomous agents is when you don't really tell them what to do."

That framing matters for anyone building autonomous systems. Most AI training today works the opposite way: engineers predefine the task, the optimization target and the metric of progress. MIT's island tested what happens when all three are stripped out.

Was it really play?

The researchers didn't just watch the agents and call it play. They compared Eko's behavior against established markers of play in animals, humans included. Those markers, as the study frames them, include:

• creating their own rules;
• repeating behaviors with slight variations;
• setting goals without external prompts;
• transforming objects and events through make-believe.

Eko checked the boxes. "Play is really about curiosity or novelty-seeking behavior, and this is essentially what Eko does," Daniela Rus, an MIT computer scientist and another coauthor of the study, told Gizmodo.

The report's title asks a question; the researchers' data answers it, effectively, with a yes.

Why does this matter for machine learning?

Beyond the entertainment value of lobster-shaped agents holding tool funerals, the study carries a substantive claim: in some cases, it may be more effective to let AI agents teach themselves through exploration and self-determination rather than through predefined tasks.

The report's conclusion puts the argument in unusually plain terms: "We normally develop artificial agents by deciding in advance what they should practice, what they should optimize, and what counts as progress. Play suggests another regime: give an agent a sufficiently rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from the history that follows. Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood."

That last line is the one worth watching. If self-directed exploration can accumulate useful knowledge — the way Eko taught itself movement, construction and even ritual on an empty island — it points to a training regime that competes with task-driven optimization for at least some capabilities.

The full behaviors of all 13 Ekos are documented in the researchers' report, published on arXiv. The question the study leaves open, and the one with commercial stakes, is which capabilities turn out to need tasks — and which ones need something closer to a childhood.

By Grace Kim