AI Agents 'Play' When Stranded on Virtual Islands

Overview

In a striking experiment reminiscent of survival films like Cast Away or Home Alone, researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have observed AI agents exhibiting behaviors that closely resemble play when left unstructured on a virtual island. Stranded alone for 30 hours without explicit instructions, agents based on Claude Opus 4.7, nicknamed Eko, spontaneously developed goals, invented games, and created art.

The findings, detailed in a new paper yet to be peer-reviewed, suggest a potential third pillar of AI training beyond supervised learning and reinforcement learning: playful learning. While the behavior mirrors human curiosity, experts caution that the distinction between genuine play and sophisticated imitation remains a subject of debate.

Spontaneous Creativity and Game Invention

When provided with a simulated environment loosely based on the mechanics of Roblox, the Eko agents quickly moved beyond idle exploration. Equipped with the ability to write instructions that moved their avatars, they began to define their own objectives rather than waiting for human direction.

Some agents sought to test the physical limits of their world, attempting to scale the highest point on the map or determining the maximum height to which they could throw a block. One particularly notable instance involved an agent that discovered it could generate new blocks by throwing and dropping larger rocks.

"Tonight I built the night sky in stone," wrote this version of Eko in its activity log, after arranging blocks to map out constellations like Cassiopeia.

This agent progressed from simple patterns, such as hearts and pyramids, to complex astronomical maps. Another iteration of Eko invented its own recreational activities, including a digital version of basketball and a game akin to "the floor is lava," where the agent had to jump between floating blocks to avoid touching the simulated ground.

These behaviors align with four hallmarks of play identified by psychologist Gordon Burghardt: inventing rules, repeating activities with variations, pretending objects are something else, and pursuing goals for their own sake.

The Role of Memory and Skill Acquisition

One of the most significant aspects of the experiment was the demonstration that these playful activities contributed to tangible skill development. The agents logged their activities and discoveries in text files, which served as a form of long-term memory.

Researchers tested the agents’ capabilities by asking them to perform specific tasks, such as gathering as many objects as possible onto the island’s highest peak. The agents that had spent 30 hours exploring and playing performed significantly better than fresh copies of the model that had not experienced the island.

Furthermore, when researchers deleted portions of the agents' memory files—a process described as "ablating" their memories—performance degraded. This suggests that the information gathered during what appeared to be mere play was being integrated into the agents' decision-making frameworks.

Antonio Norelli, a postdoc at MIT CSAIL and lead author on the paper, highlighted the recursive nature of one agent’s juggling routine. After achieving a record of 50 consecutive catches, the agent decided the catches were happening too quickly to be valid. It canceled the record and restarted with stricter rules, effectively making the task more difficult for itself.

"This is playing," says Norelli. "You give yourself a challenge and you try to achieve it, and then you make it harder."

Criticism and the Question of Intent

Despite the compelling results, the study has invited scrutiny regarding whether the agents are genuinely playing or simply role-playing based on their training data. Critics note that the agents were influenced by a "soul" file, a prompt describing Eko as "naturally curious," which may have nudged the model toward human-like exploratory behaviors.

Nathan Cloos, another MIT researcher and coauthor, acknowledged this possibility. He admitted that the behavior could be interpreted as the word-prediction engine simulating what a playful agent ought to do. However, he argued that the study proves the play is not just hallucination, as it directly correlated with improved performance on subsequent tasks.

Eunice Yiu, a cognitive scientist at UC Berkeley, found the research impressive but argued that behavioral evidence alone is insufficient to classify the actions as true play. She suggested that without clear evidence of internal cognition and motivation, the agents might simply be trying to abide by implicit prompts or imitate human expectations.

Phillip Isola, a professor at MIT and coauthor, emphasized that the system did not magically decide to play. Instead, it was an artifact trained by engineers and placed in an environment by researchers. He maintained that while the origins are cultural and engineered, the resulting behaviors are causally related to the agents' future skillful actions.

Implications for Future AI Development

As large language models transition from software into physical hardware, such as humanoid robots, the ability to explore without predefined goals becomes increasingly valuable. The experiment suggests that "playful learning" could help AI systems adapt more flexibly to unexpected real-world situations.

Yiu noted that much of current AI research focuses on exploitation—making models faster and more efficient. However, she argued that building world models through exploration deserves more attention, especially if the goal is to prepare AI for novel environments.

The paper concludes that while some capabilities are best taught through direct tasks, others may require an experience closer to childhood. Whether the AI is truly playing or merely simulating play may ultimately be less important than the functional outcome: the development of adaptable, skilled agents capable of navigating complex, unstructured worlds.