Scientists Left AI Alone on a Deserted Island With Nothing to Do. It Invented Games and Started Behaving Like Children at Play

With no assigned goals, AI agents created their own challenges — and learned from them.

21 Oct 2026, 20:28 by Tibi Puiu · ZME Science

Illustration made with the help of AI. Credit: ZME Science

What happens when you put an AI on a virtual deserted island, all alone with no task or objective? Apparently, it finds things to do by itself.

MIT researchers were amazed to see a lone figure carrying a ring to the top of a ridge, place it on the ground like a basketball hoop, and begin taking shots with a block.

The figure was an artificial intelligence called Eko, operating a virtual avatar through written commands that it gave. Nobody had told it to play basketball. In fact, nobody had given it any particular task at all.

But Eko was finding plenty of ways to entertain itself.

It invented bowling, built elaborate sculptures, and experimented with the strange physics of its little world. At one point, it discovered that throwing a particular rock hard enough caused new building blocks to appear, a hidden feature the researchers had programmed into the simulation. Eko exploited the discovery to create increasingly elaborate arrangements, eventually recreating constellations on the ground.

“Tonight I built the night sky in stone,” it wrote in its activity log.

These curious antics were part of an experiment at MIT exploring the rather wild idea whether an artificial intelligence can learn through play, much as children and animals do.

In a new study, researchers placed 13 independent AI agents on the same simulated island for 30 hours each. Without assignments, scores or rewards, the agents invented their own challenges, investigated their surroundings and developed skills that later made them better at solving problems.

Get smarter every day...

Stay ahead with ZME Science and subscribe.

Daily Newsletter
The science you need to know, every weekday.

Weekly Newsletter
A week in science, all in one place. Sends every Sunday.
No spam, ever. Unsubscribe anytime. Review our Privacy Policy.

Thank you! One more thing...

Please check your inbox and confirm your subscription.

The researchers conclude that aimless exploration might offer a powerful new way to teach AI, although whether the machines were truly ‘playing’ remains an open question.

Inventing games and defeating AI boredom

The island was sparse, comprising a few raised platforms, 12 cubes, a rock and a ring. Eko could move, jump, pick up objects and throw them. It never saw pictures, only textual descriptions of coordinates and objects.

One agent positioned the ring as a basketball hoop. Another arranged cubes like bowling pins. After hurling its ring into the ocean, Eko built a heart-shaped memorial. A different agent canceled its own 50-catch juggling record after realizing the objects had barely left its grasp, then tried again under stricter rules.

What on earth is going on?

The researchers at MIT believe the AI agents are essentially playing. They base their judgment on four hallmarks of animal and human play, which the AI also exhibited. These are inventing rules, repeating activities with variations, pretending objects were something else and pursuing activities without an outside payoff.

“This is playing,” Antonio Norelli, an MIT researcher and coauthor, told MIT Technology Review. “You give yourself a challenge and you try to achieve it, and then you make it harder.”

The real test, however, was what happened afterward.

To find out whether all this playing had taught the agents anything useful, the researchers gave them actual jobs. They compared the 13 agents that had spent 30 hours exploring the island with 13 fresh copies that had never been there before.

One challenge was to get as many objects as possible onto the island’s highest peak within an hour. The experienced agents managed an average of 10.3 objects, while the newcomers managed just 3.5. The experienced agents also generally performed better at other tasks, including building multiple towers out of blocks.

But how had they learned? The AI itself hadn’t been retrained. Instead, the agents had spent their time on the island keeping detailed notes about what worked, what failed, and which tricks they had discovered. These notes survived even when the simulation reset, allowing the agents to build on their previous experiences.

To test whether those logged activities really amounted to anything, the researchers deliberately erased specific discoveries from the agents’ notes. For instance, one agent had learned that throwing a rock could generate extra building blocks. When researchers deleted that knowledge, the agent went from building an average of six towers to just two. The experiment showed that Eko wasn’t merely acting playful. It was acquiring practical knowledge that helped it solve new problems.

There was a catch, however. The agents sometimes learned the wrong lessons. One had mistakenly concluded that stacking blocks directly didn’t work well. When researchers erased that false belief from its memory, it actually became better at building towers. Like a student who had misunderstood a lesson, Eko could carry its mistakes into future challenges.

AI-Powered Toys Are Flooding the Market and Researchers Warn the Risks Are Far Bigger Than Parents Realize
Chat GPT’s new O1 model escaped its environment to complete “impossible” hacking task — should we be concerned?
Humans and computers can be fooled by the same tricky images
Human pilots face an AI-controlled jet for the first time in a real-world dogfight

Was any of it really play?

Other projects have pursued open-ended learning, including an agent named Voyager, which learned skills in Minecraft, and Google DeepMind’s XLand. But they used automatically generated objectives or rewards. Eko’s activities arose without a specific task or scoring system telling it what to practice.

Yet the behavior was not entirely unprompted. Its initial instructions described it as naturally curious and fond of exploration, and automated messages repeatedly urged it to continue. Five agents deprived of that persona still displayed a comparable range of activities, although trials using two other AI systems produced narrower behavior.

“It could be the case that the models are genuinely playing, but we don’t know based on the paper,” Eunice Yiu, a cognitive scientist at the University of California, Berkeley, told MIT Technology Review. She cautioned that the agents might simply be imitating human play, perhaps due to some subtle conditioning in the agent’s system prompt.

And it’s not just play. The experiments reported some useful learning. For instance, some agents discovered that jumping before throwing could launch blocks far higher, a trick even the simulator’s designers had overlooked.

The approach could help robots cope with unfamiliar environments. But researchers warn that autonomous exploration belongs in controlled playgrounds, not uncontrolled real-world settings.