Strand AI on a desert island, and it starts playing and inventing games
AI agents tend to do things when we give them specific commands, such as analyzing data and answering questions. But what would they do if left to their own devices?
edited by Gaby Clark, reviewed by Robert Egan
This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility:
Add to Preferred Sources Eko's life on the island. Credit: arXiv (2026). DOI: 10.48550/arxiv.2610.07130 AI agents tend to do things when we give them specific commands, such as analyzing data and answering questions. But what would they do if left to their own devices?
That's what researchers from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) wanted to find out when they stranded several agents on a virtual desert island and left them without any specific tasks to complete.
There were 13 agents in total, all named Eko, which were built using the AI coding assistant Claude Code. Each was given a virtual body and text files to record their memories and discoveries and left alone on their own copy of Mesa World. This was a simulated 3D island built on a custom game engine called Clawblox.
The agents stayed on their islands for 30 hours, during which time they received no specific tasks or directions from humans about what activities to pursue. They were simply instructed to be curious and eager to explore.
They could not visualize the island, so instead received text descriptions of where objects were and controlled their virtual bodies by typing basic commands like jump, pick up or throw.
To keep the AI from running out of short-term memory, the island was restored to its original state every hour, and the AI's current session was reset. However, the agents did not lose their recorded knowledge and discoveries, because they could write notes, guides and discoveries into text files that survived each hourly reset.
While the agents roamed their virtual world, the researchers recorded and analyzed their behavior. What they observed is detailed in a new paper posted on the arXiv preprint server.
Instead of freezing or repeating the same basic action over and over, which you might expect when an AI isn't given any specific tasks, the agents were surprisingly creative.
Without any human prompting, they built complex towers, arranged geometric ground patterns and even improvised their own versions of bowling, basketball and juggling. In other words, they were displaying behaviors that resembled play.
The researchers reached this conclusion after evaluating the agents' actions against four classic signatures of play. These are inventing their own rules, repeating activities with variations, engaging in make-believe and pursuing goals that appeared to be for their own sake.
For instance, the AI created its own arbitrary challenges, like hopping across blocks without touching the ground, and transformed simple props, using a ring as a basketball hoop.
But this playtime wasn't just for fun. It helped them learn. Agents that had spent 30 hours on their island generally performed better on later tasks than newbie castaways. They learned techniques for building towers and placing objects on high platforms and even uncovered physics tricks that the creators hadn't realized their simulator allowed.
"The agents' self-initiated exploration thus produced not only knowledge we had hidden for them to find, but knowledge of the world that its designers lacked," wrote the study authors.
Therefore, giving AI room to play may be one way future systems learn. "Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood."
Written for you by our author Paul Arnold, edited by Gaby Clark, and fact-checked and reviewed by Robert Egan —this article is the result of careful human work. We rely on readers like you to keep independent science journalism alive. If this reporting matters to you, please consider a donation (especially monthly). You'll get an ad-free account as a thank-you.
Nathan Cloos et al., Is this machine playing?, arXiv (2026). DOI: 10.48550/arxiv.2610.07130 Journal information: arXiv
BSc Biology from University of London. BBC documentary producer with world travel experience. Freelances from southern Spain. Full profile →
MA in English, copy editor since 2021 with experience in higher education and health content. Dedicated to trustworthy science news. Full profile →
Bachelor's in mathematical biology, Master's in creative writing. Well-traveled with unique perspectives on science and language. Full profile →