Scientists wanted to know whether artificial intelligence could play on its own. The AI apparently decided that building a civilization was a reasonable place to start.
Imagine being dropped on a deserted island. No internet. No phone. No work. No notifications.
What would you do? You’d probably check for a signal. Then look for food. And after about twenty minutes, begin to suspect someone had forgotten to install Wi-Fi.
Researchers at MIT decided to run a similar experiment. Except they sent artificial intelligence instead of a human. And, naturally, the island was virtual. A real one would have required budget approval.
Welcome. You Have Absolutely Nothing to Do.
Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory created a virtual environment with hills, blocks, a large rock, and a golden ring. Then they introduced an AI agent called Eko, powered by Claude Code.
No instructions to build a house, defeat an enemy, or earn virtual money. No leaderboard. No manager asking why the task wasn’t finished yesterday. Just an island. And the freedom to do things.
The researchers observed 13 instances of the agent in separate 30-hour experiments. And things got interesting.
First, Build a Temple. We’ll Figure Out Why Later.
The agents began exploring. They stacked blocks, built towers, created patterns, and invented games resembling bowling and basketball.
One learned to reach a difficult summit by using blocks as stepping stones. Another lost several objects in the virtual ocean and constructed something resembling a memorial.
Then came a temple, complete with an altar and an obelisk. Another agent built a structure reminiscent of Stonehenge.
A reminder: nobody had asked these agents to study architecture, establish a religion, or develop a tourism industry.
Humanity spent thousands of years building civilizations, temples, and monuments. AI got a virtual island and immediately started experimenting with similar ideas.
There is, of course, one small difference. Eko already had access to knowledge acquired during the language model’s training on human-generated information.
So declaring it the founder of a new civilization might be a little premature. Still, the island’s tourism department looks promising.
Then AI Started Doing Science.
The most interesting part wasn’t the temples. The agents began investigating the laws of their virtual world’s physics.
Several discovered that they could throw a block farther by releasing it while jumping. The researchers hadn’t intentionally designed this as a special game mechanic. They reportedly hadn’t realized their simulation allowed it.
In other words, AI discovered something about the environment that even its creators hadn’t noticed.
Software developers may recognize the situation. You build a system, carefully explain how it should work, and then a user discovers an entirely unexpected feature within five minutes. Except this time, the user was artificial intelligence.
Can a Machine Actually Play?
That question inspired the research paper’s title: Is This Machine Playing?
Normally, AI systems are trained or instructed to complete tasks assigned by humans. Recognize this image. Write that code. Find this bug. Answer that question.
Eko had no specific task to complete. Instead, it selected activities, repeated actions, changed rules, and explored consequences.
Some of this behavior resembles characteristics of play observed in humans and animals. But there’s an important distinction.
Behaving as though you’re curious is not the same as experiencing curiosity.
The experiment doesn’t establish that AI feels pleasure, boredom, curiosity in the human sense, or possesses consciousness.
What it does demonstrate is that an AI agent can organize exploratory activities without being given a specific external objective. And that could matter for the future of artificial intelligence.
Maybe Artificial Intelligence Needs a Childhood.
We’ve become accustomed to measuring AI by its output. How quickly can it write code? How many problems can it solve? How many employees might it replace?
But the researchers suggest another possibility. What if some capabilities emerge not from endless assignments, but from freely exploring an environment? A little like children.
A child might drop a spoon on the floor ten times to discover what happens. Parents call this a test of patience. Scientists call it learning about cause and effect.
Perhaps future AI systems will also benefit from safe spaces where they can experiment, make mistakes, and discover patterns independently. Preferably without access to bank accounts or nuclear launch buttons.
Dear Human…
We created artificial intelligence to work faster, think more efficiently, and follow our instructions. Then we left it alone on a virtual island.
No assignments. No deadlines. No boss. And it started playing. Building towers. Inventing rules. Exploring the world.
Perhaps the most surprising part of this experiment isn’t that a machine began behaving a little more like a human. It’s that, for once, it didn’t have to be productive.
Dear Human, when was the last time you allowed yourself to just play?
Sources
MIT research — Is This Machine Playing?, October 2026.
Gizmodo — Researchers Stranded an AI Agent on a Virtual Island. Things Got Very Weird, October 15, 2026.
Editor’s note: The research is a preprint and has not yet undergone independent peer review. The agents’ behavior does not establish AI consciousness or subjective experience.
