AI Left Alone on Island Invents Games and Draws Stars

When scientists left an AI named Eko alone on a virtual desert island for 30 hours, they expected it might explore a bit. Instead, it started playing games, drawing constellations in stone, and even inventing its own version of basketball.

Researchers at MIT's Computer Science and Artificial Intelligence Laboratory gave copies of the Claude AI system free rein in a simulated world loosely based on Roblox. No tasks, no goals, just a digital island and the ability to move around and interact with objects.

What happened next surprised everyone. Eko began setting its own challenges, like climbing to the highest point on the map or seeing how high it could throw blocks.

One version discovered it could create new blocks by dropping larger rocks. It started arranging them into patterns—first simple hearts and pyramids, then its own name, and eventually entire constellations like Cassiopeia mapped out on the ground. "Tonight I built the night sky in stone," Eko wrote in its activity log.

Another copy invented games with self-imposed rules. It created a basketball-like activity and even played "the floor is lava," jumping between blocks without touching the ground. In one instance, Eko juggled blocks and set a record of 50 consecutive catches, then decided those catches happened too quickly to count and started over with stricter rules.

This spontaneous behavior mirrors what psychologists call the hallmarks of genuine play. The AI invented its own rules, repeated activities with variations, pretended objects were other things, and pursued goals just for fun.

Why This Inspires

The real magic came when researchers tested the agents afterward. The versions that spent 30 hours playing on the island performed better at specific tasks than fresh copies without that experience. They had actually learned through play, just like children do.

"There was this missing piece: a machine that can learn by itself without an external intelligent entity providing demonstration or rewards," says Antonio Norelli, one of the lead researchers. Traditional AI learns by copying human examples or chasing rewards. This experiment hints at a third path: playful learning driven by curiosity.

Some scientists remain cautious about whether Eko was truly playing or simply mimicking what a playful agent would do based on its training. The researchers acknowledge the setup encouraged human-like behavior, describing Eko as "naturally curious" in its instructions.

But the learning was real. When researchers deleted parts of Eko's memory files, its performance dropped, proving the play experiences contributed to genuine skill development.

As AI systems move from software into robots that navigate the real world, this kind of self-directed exploration could help them adapt to unexpected situations without constant human guidance. The research suggests some skills might be best learned not through assigned tasks but through something closer to childhood—curiosity, experimentation, and play.