What happens if you give artificial intelligence freedom of action without setting any tasks for it? Researchers from the Massachusetts Institute of Technology Computer Science and Artificial Intelligence Laboratory (MIT CSAIL) tried to find the answer to this question. They placed several AI agents on a virtual desert island and left them without specific instructions.
The experiment involved 13 agents, each of which bore the name Eko. All of them were created using an AI-based programming assistant called Claude Code. Each agent was given a virtual body and text files for recording memories and discoveries. They were then placed one by one on their own copy of Mesa World, a simulated three-dimensional island built on the Clawblox game engine.
The agents spent 30 hours on the islands. During this time they received no specific tasks or instructions from humans about what to do. The only instruction was to show curiosity and strive for exploration. Since the agents could not see the island, they received text descriptions of the locations of objects and controlled their virtual bodies by entering simple commands such as "jump", "pick up" or "throw".
So that the AI would not exhaust its short-term memory, every hour the island was returned to its initial state, and the agent's current session was reset. However, the agents did not lose the knowledge and discoveries they had recorded, since they could save notes, guides and findings in text files that survived the hourly reset.
Observing the behavior of the agents, the researchers recorded and analyzed their actions. The results are described in a new paper published on the arXiv preprint server.
Instead of freezing or repeating the same simple action, as might have been expected in the absence of tasks, the agents showed surprising creativity. Without any hints from humans, they built complex towers, laid out geometric patterns on the ground and even improvised their own versions of bowling, basketball and juggling. In other words, their behavior resembled play.
The researchers reached this conclusion by assessing the agents' actions against four classic signs of play: inventing their own rules, repeating actions with variations, engaging in pretend activity and pursuing goals that apparently are achieved for their own sake. For example, the AI created arbitrary challenges, such as jumping from block to block without touching the ground, and turned simple objects into play attributes, using a ring as a basketball hoop.
However, this play was not just entertainment. It helped the agents learn. Those who spent 30 hours on the island generally coped better with subsequent tasks than newcomers. They mastered techniques for building towers and placing objects on high platforms and even discovered physical tricks that the creators of the simulator had not suspected.
"The agents' independent exploration produced not only the knowledge we hid for them, but also knowledge about the world that its creators did not have," the authors of the study write.
Thus, giving AI the opportunity to play may become one of the ways to train future systems. "Perhaps some abilities are best conveyed through tasks. Others may need something more, resembling a childhood," the researchers conclude.