Artificial intelligence agents given no assignment on a simulated desert island began creating challenges for themselves, and researchers at MIT say the behavior has the hallmarks of play. The findings appear in a paper posted on arXiv, a site where studies are shared before peer review.
The team placed 13 AI agents, each called Eko and running on Claude Opus 4.7, on the island by themselves for 30 hours. Afterward, the agents were given specific tasks, such as getting as many objects as possible onto the island's tallest rock spire. According to the paper, the Eko agents that had spent those 30 hours playing generally outperformed identical agents that had never been on the island.
The researchers see the approach as an alternative to current practice, in which developers decide beforehand what an AI should practice and how success is measured. They suggest it could be a different route to teaching skills, especially ones that might eventually be used to control robots. "Play suggests another regime: give an agent a sufficiently rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from the history that follows," they write.
One agent's log illustrates the pattern. In its first hour it wrote, "Let me observe my surroundings," then set out to conquer the tallest spire. After failing, it returned an hour later with a backup plan: it hauled a rock up a neighboring summit, stacked two blocks on top and jumped. "I'M ON THE SPIRE," it wrote.
Its next big moment went less well. While testing how far it could throw a golden ring, it sent the ring into the ocean. "Ouch. Lost the Ring," it wrote, before deciding to "pivot to something achievable and meaningful" by building a heart of stones as "a memorial to the lost Ring."
When the game reset and the ring reappeared, the agent used it as a basketball hoop, a target and eventually an accessory. The game would not let it put the ring around its own body, so it found a workaround: falling off a cliff edge would make it respawn in the same area as the ring. It stepped off and dropped through the middle of the ring. "I'm wearing the Ring," it announced.
The researchers say such stunts were more than curiosities. Six of the 13 agents learned they could throw a block farther by releasing it mid-jump, something the researchers had not intended to make possible. The exploration, they wrote, produced "knowledge of the world that its designers lacked."
The setup did include some nudges toward humanlike behavior. Eko was never told to invent games, but its master prompt, a "soul" file, described it as "naturally curious."
Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Technology Review she was impressed but wants more evidence of the cognition behind the behavior. "It could be the case that the models are genuinely playing, but we don't know based on the paper," she said.
The researchers conclude that different methods may suit different capabilities. "Perhaps some capabilities are best taught by tasks," their report says. "Others may need something closer to a childhood."
