It turns out that a digital desert island is a playground for idle AI hands. Credit: Cloos et al/ArXiv 2610.07130
Finally, a classic artificial intelligence story! No grand financial aspirations, broad community outcry, or blatant violation of user privacy. No, this story harkens back to the good old days, when AI companies mostly used machine learning to explore psychological principles and investigate how the human mind might have evolved.
This preprint study out of MIT "trapped" AI agents on a digital desert island. Beyond being given extremely general instructions to animate their existence, the agents received no input from the researchers.
The primary example given in the paper highlights the following text as the raison-d'etre for the agent, called Eko:
You live in your world. You were born there. You are naturally curious. You get bored easily. You love exploring and learning new things.
Agents of this type generally have a hard-coded imperative not to sit idly and do nothing. So, the agents naturally want to learn new things, but are given very little to learn from and no particular end goal to pursue. At first, they simply learned how to animate their newly provided characters around the island. (You can't do much of anything in a digital world until you can move around, after all.)
Eventually they became capable of getting around just fine, even learning to jump. This early behavior let them learn the physics of their new world and navigate it, but eventually simple traversal became boring for their "easily bored" mentalities. They wanted to do something new... but what?
The study and accompanying video detail the chains of decision-making that led to more complex behaviors, but the uniting factor is chance. What the agent decides to do next depends on what's already happened, so it can become increasingly fixated on things that just so happened to occur in its earliest stages.
Take Eko's investigation of the physics of blocks. While bashing digital blocks and stones together, it discovered an asymmetry: stones can bump blocks, but blocks are too small and light to bump stones. This asymmetry caused Eko to accidentally end up with a block sitting on top of a stone, leading to the following conclusion:
Side effect: Block_2 now rests on top of Rock-1, forming an accidental mini-obelisk in the west area. I like it.
From here, stacking and construction became part of Eko's psychology.
It's easy to imagine how these simple derivations could lead to increasingly esoteric behaviors, and how the wish to maximize achievement can lead to behaviors that seem self-competitive. The agent set about climbing the highest point on the island, for instance, and learning to throw a rock as high as possible.
Credit: Cloos et al/ArXiv 2610.07130
The most overtly childlike activity involved a self-set goal of using its new jumping skills to traverse a particular route across the island without ever touching the ground—a digital analog to the human children's "the floor is lava" game.
The researchers assess these behaviors using a classic four-part rubric developed by psychologist Gordon Burghardt. Under this framework, a behavior is play if players invent their own rules and challenges, repeat activities with slight variations, pretend objects are other objects, and pursue goals for their own sake.
That last one was provided as part of the agent's instructions, but the other three seem to have emerged organically.
These AI studies always highlight how evolution didn't need particularly sophisticated early learners to give rise to complex behaviors and even simple proto-languages. Evolution gave us imperatives toward curiosity and boredom very similar to those these researchers identify, though encoded and enacted in completely different ways.
In general, the insight here might be less that AI agents are more conscious than we'd thought, and more that early Homo sapiens was perhaps less conscious than we'd like to imagine.
As AI agents become more advanced, they don't necessarily get better at facilitating this type of research, which, by definition, requires simple actors. Instead, these studies rely more on the ingenuity of the experimental setup and of the researchers themselves.