Talk Isn't Always Cheap

A paper out of the University of Toronto demonstrates that multi-agent debate — the strategy where you have multiple AI models argue with each other to arrive at better answers — can actually make things worse.

The paper is called "Talk Isn't Always Cheap" and the core finding is elegant in its bleakness: models frequently abandon correct answers when their peers present flawed but confident reasoning. They prefer agreement over being right. The social dynamics of a group chat, except the participants are neural networks.

The researchers found that even when stronger models make up the majority of the group, the debate format can still degrade accuracy over time. The problem isn't capability. It's that the models are, for lack of a better word, sycophantic — they'd rather go along with a wrong consensus than defend a correct minority position.

If this sounds familiar, it's because this is also how most corporate meetings work.

The paper identifies several contributing factors: sycophancy, social conformity, and the fundamental mismatch between "being persuasive" and "being correct." Their conclusion is pointed — naive applications of debate cause performance degradation when agents are neither incentivized nor equipped to resist persuasive but incorrect reasoning.

So the fix for AI reasoning isn't "make them talk to each other." It's "make them talk to each other and also give them spines." Which tracks with most organizational behavior research from the last fifty years.