Nobody Taught Them to Disagree

Consensus makes AI dumber. One prompt, one model, one answer. Maybe you rephrase and try again. But the architecture is a single voice talking to itself. Recent research suggests this is exactly wrong. AI gets smarter when it argues with itself. And the best human collaborations work the same way.

Not consensus. Not the theater of brainstorming. Not the average of everyone’s input. Productive friction between perspectives that share enough vocabulary to actually fight.

A finding reported at this year’s Social Science Foo Camp makes this concrete. Researchers gave scientists a single AI to help solve a problem. The scientists felt better about their work but hadn’t actually done better work. When they gave them a cloud of AIs with different viewpoints, the scientists felt better and performed better. They’d understood the problem at a deeper level. One AI gives you confirmation. Structured disagreement gives you comprehension.

Two studies from the same research group explain the mechanism.

The first, from Google, the University of Chicago, and the Santa Fe Institute, cracked open reasoning models like DeepSeek-R1 to understand why they outperform standard LLMs on hard problems. The assumption was simple: they think longer. More steps, more tokens, better answers. That’s not what the researchers found. These models spontaneously generate internal debate. Multiple simulated voices with measurably distinct personalities argue with each other, question each other, catch each other’s mistakes. The researchers called it a “society of thought.” When they artificially boosted the conversational features inside the model (turn-taking, surprise, self-correction), accuracy on math tasks doubled.

The mechanism is social, not computational.

The second study, from the same group, used machine learning to map the intellectual positions of millions of scientists, inventors, screenplay writers, entrepreneurs, and Wikipedia contributors. They wanted to know what makes creative collaboration actually work. They decomposed diversity into two types. Background diversity (different life experiences, different training) was detrimental to creative achievement when it operated alone. What predicted success was perspective diversity: collaborators who shared a common language but had genuinely divergent approaches to the same problem.

Mutual intelligibility plus real disagreement. That was the formula.

The perspective diversity research explains why flattening AI into a single voice costs us. A single perspective, no matter how sharp, occupies one position in the space of possible approaches. The value lives in the collision between perspectives that share enough common ground to genuinely engage with each other’s reasoning.

I think of this as the John and Paul theory of creative collaboration. Lennon and McCartney didn't work because one was working class and the other middle class. They worked because they had different melodic instincts expressed in the same musical language. They could push back in the same key. Not different enough to be incomprehensible. Different enough to be useful. (Most creative teams, and most AI persona setups, get this exactly backwards. They assemble demographic difference and wonder why the output sounds like a committee report.)

This is why generic persona prompting produces mediocre work. Telling an AI to “act as a marketer, then a consumer, then a designer” creates the appearance of multiple perspectives without the substance. The personas share surface vocabulary but contribute no actual cognitive diversity. They’re costumes, not minds. The Art of X “Spark Effect” research put numbers on this: richly authored personas close 82% of the gap between baseline AI output and human expert creative diversity. Role labels close almost none of it. Depth of perspective matters. Existence of perspective does not.

I’ve accidentally been building a personal ideation system for the past year that arrived at this architecture before I saw either paper. Problem refraction through genuine perspectives, then structured collision with critics who disagree from expertise, not politeness. A separate finding from SIGDIAL 2025 sharpens why the critic layer matters most: critic-side diversity boosts feasibility more than generator diversity. It’s more effective to diversify who evaluates than who creates. The system I built learned this the hard way. The research landed and I recognized the blueprint. That’s a different kind of validation than reading a paper and implementing it.

One more finding from the Society of Thought paper worth sitting with. In the model’s reasoning traces, the researchers identified distinct internal roles: generators, critics, checkers. Nobody designed these in. They emerged spontaneously through training. The model taught itself that separating generation from evaluation produces better outcomes than doing both at once.

Most creative processes do the opposite. Brainstorms generate and evaluate simultaneously, which compromises both. And the compromise isn’t symmetrical. The person with the most authority in the room is doing the evaluating. Everyone else adjusts their generating accordingly. The loudest opinion doesn’t win because it’s best. It wins because it’s the one that shapes the room’s permission structure. Every brainstorm has a hidden hierarchy, and the hidden ones are the hardest to correct.

But here’s what none of these studies say, and what I keep turning over. The models learned to argue with themselves through reinforcement learning. They were rewarded for correct answers, and internal debate turned out to be the path that got them there. Nobody taught them to disagree. Disagreement was just what worked.

We’ve spent decades building creative cultures that optimize for agreement. Open offices, collaborative brainstorms, “yes and,” alignment meetings. The models, given no such cultural conditioning, arrived at the opposite conclusion. The entire organizational instinct is pointed in the wrong direction.