AI That Copies Its Peers Saves Effort — But Isn't Smarter

Modern AI models can already rewrite the instructions they follow, essentially teaching themselves new tricks. This study asked a different question: if each AI is chasing its own goal, and can peek at what its peers are doing, does the whole group get smarter? The researchers gave three AI models a shared budget of "tokens" — think of these as the metered units of AI effort, like minutes on a phone plan — and let them write, revise, and borrow written skill files.

Copying clearly helped with efficiency. Watching peers changed how the models improved: one found useful skills sooner, and another spent less of its budget on private trial and error. But here's the surprise — neither model ended up performing better than a model working alone with the exact same budget. Some models explored too narrowly, and others burned through their tokens before they ever got around to doing the task.

The bigger catch was sameness. Skills got copied, tweaked, and passed along, so one good discovery could spark more searching. But those exchanges pulled the whole population toward fewer independent discoveries. In other words, the group converged on the same few tricks and stopped finding new ones.

Why should you care? This is early lab research, not something running inside your chatbot tomorrow. But companies are already building teams of AI "agents" (AI that can take actions for you) to handle research, sales, and coding. If those teams mostly copy each other, you may get cheaper answers rather than better ones — and everyone sharing the same blind spots. It's like an office where everyone reuses the same spreadsheet template: faster, sure, but nobody catches the error hiding in row 40.