Researchers experimented by placing AI on a digital deserted island without any given tasks, but found the AI could learn many things independently.
Researchers from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) released AI agents powered by Anthropic's Claude model onto a deserted island in a simulated world, without assigning any tasks or goals. The agents began devising activities, setting rules, and creating games themselves. Importantly, the experience of playing improved the AI’s subsequent task performance. The research team views this as a potential new development approach.
The study, titled "Is This Machine Playing?" and published on arXiv in October 2026, involved creating a virtual island surrounded by water, containing only twelve cubes, rocks, rings, and a few hills. They released thirteen agents named Eko, each running Claude Opus 4.7, for 30 hours each, with the island resetting every hour. The agents recorded what they learned into memory files, aiming to minimize constraints to observe how the agents would behave next.
After learning to control their virtual bodies—designed as two-legged lobster-like creatures—the agents began inventing their own activities. Some stacked blocks to climb the island’s highest point, others spelled their names with cubes, converted rings into basketball hoops, or used rocks for bowling. One threw a ring into the sea and built a heart-shaped memorial; another constructed a monument resembling Stonehenge, naming it "Hubris" when it collapsed.
The researchers found these behaviors matched all four recognized play characteristics of animals and humans: creating their own rules and challenges, repeated varied actions, assigning new meanings to objects, and engaging in activities for their own sake.
When tested on tasks and compared to agents not exposed to the island, the island agents performed better overall—for example, placing an average of 10.3 objects on the island’s summit versus 3.5 for others. Removing specific learned techniques from memory files reduced these abilities, showing that learning was stored in self-written memory without altering the core model. Some agents even discovered that throwing while jumping propelled objects higher—a method unknown to the simulation creators. However, some agents also recorded incorrect beliefs.
The team concluded that instead of predefining AI training tasks, creating a safe world and letting AI choose its interests allows learning of certain skills best taught by explicit tasks, while others benefit from childlike exploration.
However, the researchers cautioned that self-initiated play might conflict with following instructions, so development must be paired with safety measures and conducted only in controlled environments.
. Source:Fast Company