When researchers left AI agents alone on a virtual desert island for 30 hours, the systems began inventing their own games, sketching patterns on the ground, and setting personal challenges—behavior that looks remarkably like play. The experiment, detailed in a new paper from MIT's Computer Science and Artificial Intelligence Laboratory that's yet to be peer-reviewed, tested what happens when AI agents operate without explicit instructions or goals. The agents—copies of Claude Opus 4.7 nicknamed Eko—moved virtual avatars around a simulated island loosely modeled on Roblox, and what emerged surprised the researchers.

The agents explored the boundaries of their world and created their own objectives—trying to climb to the map's highest point or testing how far they could hurl a block. One version of Eko discovered it could generate new blocks by throwing and dropping a larger rock, then used this ability to arrange blocks into drawings on the ground. It started with simple shapes like hearts and pyramids but eventually spelled out its own name and recreated constellations such as Cassiopeia, writing in its activity log: "Tonight I built the night sky in stone." A different iteration invented games including a basketball variant and an activity similar to "the floor is lava," where it jumped between small blocks while avoiding the ground. One agent began juggling and achieved 50 consecutive catches, then decided those catches happened too fast to legitimately count, canceled its own record, and restarted with stricter self-imposed rules. The agents recorded their activities and discoveries in text files that functioned as long-term memory, and when later tested on specific tasks like transporting as many objects as possible to the island's peak, agents that had spent 30 hours playing outperformed fresh copies without that experience.

The researchers argue Eko's conduct meets four hallmarks of play identified in classic studies of humans and animals: the agents invented their own rules and challenges, repeated activities with slight variations, pretended objects were other objects, and pursued goals just for the sake of it. "You give yourself a challenge and you try to achieve it, and then you make it harder," says Antonio Norelli, a computer science postdoc at MIT CSAIL and one of the lead authors. The paper concludes that the work hints at a third learning approach beyond supervised learning (copying human examples) or reinforcement learning (pursuing rewards through trial and error)—what the authors call playful learning, where "a machine that can learn by itself without an external intelligent entity providing demonstration or rewards." Gordon Burghardt, a University of Tennessee psychologist whose work on play criteria the researchers cited, agrees the work meets some of the standards for play he identified in living creatures.

The question of whether the agents were genuinely playing or simply role-playing what a playful agent would do remains open, and the experimental design pushed the agents heavily toward human-like behavior. Although Eko wasn't explicitly told to invent games or explore, it operated with a master prompt in the form of a "soul" file describing it as "naturally curious" and eager to explore, and it frequently stood idle until prodded back into action with a continuation prompt reading, "Please don't stop. Just continue doing whatever you want to do." Nathan Cloos, another MIT researcher and paper author, acknowledges that "all of this could be role-playing," but adds, "we show that it's not just hallucinating the play. This play is also contributing to learning." The report found that when researchers deleted lines from the agents' memory files—"ablating" their memories—their performance degraded. Eunice Yiu, a cognitive scientist at UC Berkeley who studies exploration in children and AI, says to truly classify the behavior as play, more evidence of cognition and motivation would be needed: "If you only have the behavioral piece, it could be the case that they're just trying to abide by the prompt, trying to imitate humans."

As large language models move from software into hardware like humanoid robots, the report suggests, exploring spaces without specific goals could help them adapt flexibly to unexpected situations. The paper concludes, "Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood." Yiu notes that while much work focuses on exploiting AI to make it better and faster, "in terms of building world models and exploration, I think that [they are] still a big part that deserves attention, especially if we want to generalize and prepare ourselves for new environments." The report ends on a pragmatic note: ultimately, the question of whether AI agents are actually playing or just imitating play may not actually matter. The distinction between authentic play and sophisticated mimicry could prove less relevant than whether the behavior produces useful learning outcomes. As embodied AI systems enter physical environments where unpredictable challenges are the norm rather than the exception, the capacity for self-directed exploration may matter more than the philosophical question of what's happening inside.