When you ask three AI systems the same question, they give you the same answer — and it costs more than asking one

Researchers tested a popular trick where you run the same AI model multiple times with different role assignments and vote on the answers. It turns out the different 'agents' almost always reach identical conclusions, meaning you're paying three times the cost for no real diversity. A cheaper approach that weights answers by how different they are reaches the same accuracy at a quarter of the token cost.

Many teams building AI systems assume that prompting the same model in different ways creates genuine disagreement — the way a real committee of different experts would. This paper measures what's actually happening in the black box and finds the different agents are thinking in nearly identical patterns. That's a structural problem: you can't improve accuracy by aggregating clones. The cheap fix they propose suggests most teams are wasting compute on fake diversity rather than building systems that actually reason in genuinely different ways.