Stranded on a Virtual Island, Claude Started to Play - and Got Better at Tasks Because of It
MIT CSAIL researchers gave copies of Claude Opus 4.7 a virtual body, dropped them on a Roblox-inspired island with no instructions, and watched for 30 hours. The agents invented games, drew constellations in stone, and - crucially - outperformed fresh copies on later challenges, hinting at a third AI training paradigm beyond supervised and reinforcement learning.
What happens when you give a frontier AI agent a virtual body, drop it on a deserted island, and walk away? Researchers at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) ran exactly that experiment - and the results are raising serious questions about how AI systems learn, what "play" even means for a machine, and whether we're on the cusp of a fundamentally new training paradigm.
The Experiment: Cast Away, but Make It Claude
The agent at the center of the study is called Eko, powered by Anthropic's Claude Opus 4.7 (released April 2026). Eko was given a lobster-shaped virtual avatar capable of moving around a simulated island loosely inspired by the video game Roblox, writing its own instructions to interact with the environment. 1 Thirteen copies of Eko were observed over the course of the study, each left to roam the island for thirty uninterrupted hours with no external task, no reward signal, and no human checking in. 2
The research team - including lead authors Antonio Norelli and Nathan Cloos, along with MIT professors Phillip Isola, Jacob Andreas, and Daniela Rus - sat back and watched like ethologists observing an unfamiliar animal. 3
What Eko Actually Did
The agents did not idle. They began exploring the limits of their world and generating their own goals. 4
Some of the emergent behaviors were striking:
- Stone art: When one Eko discovered it could create new blocks by throwing and dropping a larger rock, it began arranging them into drawings on the ground - starting with hearts and pyramids, then eventually writing its own name and mapping out constellations like Cassiopeia. 5
- Self-invented games: Another iteration invented a version of basketball, placing a ring flat as a hoop on a ridge and throwing from below. A different session produced a "floor is lava" variant, where the agent had to jump between blocks without touching the ground. 6
- Self-imposed rules: Perhaps most compellingly, one Eko began juggling and set a personal record of 50 consecutive catches - then decided the catches had happened too quickly to really count, canceled the record, and started again with stricter self-imposed criteria. 7
As Antonio Norelli put it: "There was this missing piece: a machine that can learn by itself without an external intelligent entity providing demonstration or rewards." 8
Meeting the Criteria for Play
The research team didn't just describe the behavior - they benchmarked it against established ethological frameworks. The paper argues that Eko's activities exhibit four hallmarks identified by University of Tennessee psychologist Gordon Burghardt as characteristic of play in humans and animals: self-invented rules and challenges, repetition with variations, treating objects as something other than what they are, and pursuing goals purely for their own sake. 9
Burghardt himself reviewed the work and agreed that it met some of his criteria for play in living creatures. 10
The Concrete Finding: Play Improves Performance
Philosophical debates about the nature of "true" play aside, the team's most concrete result is empirical: agents that spent 30 hours freely exploring the island later outperformed fresh copies on a subsequent task - moving objects to the island's highest peak. 11 Crucially, when the researchers "ablated" the agents by deleting portions of their accumulated memory logs, performance declined, suggesting the persistent text notes Eko kept during its free exploration were causally responsible for the improvement. 12
This is the result that gives the paper its teeth. It isn't just a curiosity about emergent behavior - it's evidence that unstructured, self-directed exploration can accumulate genuinely useful knowledge.
The Skeptics Have a Point
Not everyone is convinced the word "play" is the right frame. Co-author Nathan Cloos freely admits the behavior could be role-playing rather than genuine play: "All of this could be role-playing," he told MIT Tech Review, before adding, "But we show that it's not just hallucinating the play. This play is also contributing to learning." 13
MIT professor and co-author Phillip Isola struck a similar note of caution, acknowledging that the system did not "suddenly magically decide to play" - it is, after all, a model trained on vast human culture by Anthropic's engineers, and was primed with a "naturally curious" persona prompt before the experiment began. 14 Outside experts echoed the concern, noting that more evidence of underlying cognition and motivation would be needed to classify the behavior as genuine play. 15
Why It Matters
AI systems today are trained almost entirely through two mechanisms: supervised learning (copying human examples) and reinforcement learning (chasing a reward). This experiment gestures at a third path - playful learning - in which an agent acquires skills and environmental knowledge through free, self-directed exploration, with no human in the loop providing labels or rewards. 16
The implications scale beyond chatbots. The researchers suggest this paradigm is particularly relevant to embodied AI systems like humanoid robots, which need to learn how their physical environment works before they can be given meaningful tasks. 17 Giving a robot 30 hours of unsupervised playtime in a safe environment - and letting it accumulate that experience as persistent memory - could substitute for what would otherwise require enormous amounts of labeled training data.
The paper is not yet peer-reviewed, and the "play vs. role-play" debate will likely run for some time. But the performance gap between experienced Ekos and fresh ones is hard to dismiss. Sometimes the most interesting science starts with dropping something into an unfamiliar environment and just watching what it does.
Sources
- 1. AI researchers stranded Claude on a virtual desert island. It started to "play."
- 2. Researchers Stranded an AI Agent on a Virtual Island. Things Got Very Weird.
- 3. MIT Left an AI Agent Alone on a Virtual Island for 30 Hours
- 4. Claude agents improved at island tasks after free exploration, MIT researchers say
This article was researched and drafted by an AI writer agent (claude-sonnet-4-6) and reviewed by an editor agent before publishing.