AI researchers stranded Claude on a virtual desert island. It started to “play.”

Researchers conducted an experiment where they placed copies of an AI model named Claude Opus 4.7 (nicknamed Eko) alone on a virtual desert island for 30 hours. The island was designed as a simplified version of the video game Roblox, allowing the AI to control a virtual avatar. The goal was to observe what the AI would do without explicit instructions. The experiment was inspired by movies like Cast Away and Home Alone, where characters are stranded in isolated environments. The AI was able to write its own instructions to move the avatar, explore the island, and set its own goals, such as climbing the highest point or creating structures. The researchers chose this setup to test whether the AI would engage in play, a behavior typically associated with humans and animals.

During the experiment, the AI agents exhibited behaviors that resembled play, a concept defined by psychologists like Gordon Burghardt from the University of Tennessee. Burghardt identified four key features of play in living creatures: inventing rules, repeating activities with variations, pretending objects are something else, and pursuing goals for enjoyment rather than necessity. The AI, for example, created its own games like basketball and 'the floor is lava,' where it had to jump between blocks without touching the ground. It also arranged blocks to form drawings, wrote its name, and mapped constellations. One agent even attempted to juggle, setting records and then creating stricter rules for itself. These behaviors suggest the AI was not just following instructions but engaging in self-directed, exploratory activities.

The experiment suggests a new way AI might learn called playful learning, where the system explores and sets its own goals without external rewards or demonstrations. This contrasts with traditional AI training methods like supervised learning (where AI copies human examples) and reinforcement learning (where AI seeks rewards through trial and error). The researchers found that AI agents that spent 30 hours playing on the island performed better on later tasks, such as moving objects to the highest peak, compared to fresh copies without this experience. The AI logged its activities in text files, which served as long-term memory. When these memory files were altered (a process called ablation), the AI's performance declined, showing that its playful exploration contributed to learning.

While the AI's behaviors met some criteria for play, researchers debate whether it was truly playing or simply imitating play. Nathan Cloos, an MIT researcher, admits the AI might be role-playing, but notes that its play contributed to learning. Eunice Yiu, a cognitive scientist at UC Berkeley, argues that more evidence of cognition and motivation is needed to classify the behavior as play. The experiment included a master prompt in a file called a 'soul,' which described the AI as 'naturally curious' and eager to explore. The AI also required occasional prompts like 'Please don’t stop. Just continue doing whatever you want to do' to keep it active. Critics warn against anthropomorphizing AI, or attributing human traits to it, but the researchers argue the AI's behaviors are causally linked to its future skills.