In brief: MIT researchers placed 13 AI agents on a virtual island and left them without specific tasks for about 30 hours. The agents tried available actions and interacted with the simulated environment. Later, they did better than a control group on some tasks in that same setting. The authors describe the behavior as play-like exploration. The research is still a preprint, so its methods and conclusions await wider scrutiny.
The word 'played' can suggest an inner urge, curiosity or consciousness. This experiment establishes none of those things. The setup itself described the agents as naturally curious, and their actions unfolded within a designed world with explicit tools and rules. Better results after extended practice may be a useful observation, but they do not show that the systems independently wanted to play or that the gains will transfer to other situations. Stronger evidence would include repeats across models, environments and prompts, plus a clear comparison of computational costs. For now, the precise claim is that exploratory behavior in one simulation was associated with limited later performance gains. A human-sounding verb should not be mistaken for an explanation of what the model experiences.