Claude Just Wants to Have Fun? Without Any Work, AI Agents Like to ‘Play'

How do advanced AI agents behave when they're trapped on an island with nothing to do? According to new research , their undirected activity looks a lot like human play—and it actually helped them perform tasks better later in the experiment.

MIT researchers ran 13 separate virtual avatars, each controlled by an AI agent based on Anthropic's Claude Opus 4.7, and placed each one alone on its own copy of a simulated desert island. The researchers' environment looks broadly comparable to the popular game Roblox and uses the same programming language. The AI avatars can explore, interact with objects on the island, and even build structures.

In the beginning, the agents were not given any particular tasks to work on. Each "run" lasted 30 hours, with the island resetting every hour while the agents kept their memories.

The agents engaged in a wide range of activities that varied markedly across agents despite identical starting conditions. Agent 12, for example, spent most of its time forming shapes such as a compass or a clock, and figures such as a butterfly or a sailboat. Meanwhile, Agent 9 spent almost an entire hour spelling its name in cubes.

The agents also played simplified sports, using a ring left on the island as a target to throw cubes in from a distance or setting up a rock to knock over cubes arranged like bowling pins. In contrast, some agents did nothing. The games the agents played would change, and often become more complex over time. In one instance, a simple game of catch later evolved into variations that included multiple objects, timing rules, and even juggling.

The researchers claim the agents performed the tasks for their own sake rather than for a human-assigned purpose, yet they still worked toward a goal while following rules. "They are often organized around goals, but these goals give form to the activity rather than serving some further purpose," the researchers write.

It turns out all this goofing off makes them more effective later in the study.

The researchers pitted the agents that had played multiple rounds against an identical set of agents, though the second group lacked the playing agents' memories. The first group generally outperformed the fresh group. Six-in-ten (62%) of the 30-hour agents were able to build more than two towers, whereas only 15% of the 0-hour agents did so.

It may be a good idea to take the study with a pinch of salt. MIT Technology Review points out that although the agents were not explicitly told to play games or explore the island, they did have a master prompt in their "soul" file, a text document that described them as "naturally curious" and eager to explore.

"It could be the case that the models are genuinely playing, but we don't know based on the paper," Eunice Yiu, a cognitive scientist at UC Berkeley who studies exploration in children and AI, told MIT Technology Review. "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."

The publication also pointed out that previous research on AI agents "has been criticized for anthropomorphizing them," adding that the study "did a lot to nudge the agents toward behaving like humans."

PCMag and Yahoo may earn commission from links in this article.