The Wisdom of Inefficient Disagreement

Ask twenty-five intelligent people to describe time, and you may get twenty-five different answers. A physicist might call it a dimension. A poet might call it a thief. A parent may measure it by how quickly a child grows.

Those differences are not random. They carry traces of memory, culture, experience and the particular way each person has learned to see the world. We can look at the same mystery and somehow arrive at entirely different interpretations.

So what happens when we ask twenty-five artificial intelligences the same question?

In a recent study of language-model diversity, researchers asked twenty-five AI models to generate fifty metaphors about time. Across more than a thousand responses, the models repeatedly returned to two dominant ideas:

Time is a river.

Time is a weaver.

The researchers called the broader phenomenon an “Artificial Hivemind.” Across thousands of open-ended prompts, models built by different companies and trained in different ways produced answers that were far more similar than we might expect.

Predictably, the phrase generated headlines suggesting that AI had somehow converged into a single universal mind. The kind of dystopian framing that would make Aldous Huxley blush.

I think the sensationalism distracts from a more interesting question.

What happens when the systems we increasingly rely on to expand our thinking begin narrowing the range of ideas we encounter?

I’ve spent most of my career in rooms where people disagreed about what should be built next. Designers disagreed with engineers, engineers disagreed with project managers and clients disagreed with everyone.

At the time, I thought of much of this as unavoidable friction. Something you worked through so everyone could stop arguing and get back to building.

However, we may have misunderstood some of that friction.

Some of the best ideas I’ve been part of emerged because someone in the room refused to move on. They challenged an assumption everyone else had accepted, pushed an idea we wanted to dismiss or asked an annoying question that forced us to explain why we believed what we believed.

Sometimes they were wrong, many times they simply made the meeting longer but every so often, they changed the direction of the work.

We tend to treat disagreement as evidence that a group has failed to reach the answer. But disagreement can also reveal that the answer is less obvious than it appears.

Picasso’s Les Demoiselles d’Avignon is a useful example. When friends and fellow artists first encountered the painting in his Paris studio in 1907, many were shocked by its fractured forms and rejection of traditional perspective. Even Georges Braque, who would soon work closely with Picasso in developing Cubism, initially reacted negatively to it.

What eventually helped reshape modern art began as something many intelligent people simply could not accept.

That pattern appears everywhere. New scientific theories, artistic movements and technological shifts often begin outside the boundaries of accepted thinking. Someone looks at what everyone else considers settled and says, in one form or another, “I don’t see it that way.”

Clearly disagreement slows us down but maybe speed was never the right way to measure its value.

The AI hivemind research becomes more interesting when viewed through that lens.

The concern is not that every AI will give us exactly the same answer. They won’t. Different models use different language, examples and structures.

The more subtle problem is that different words can still carry the same underlying idea. Twenty-five models can produce hundreds of metaphors and still spend most of their time circling the same two.

Different words. Same intellectual neighborhood.

That matters because we are changing what we ask these systems to do.

AI is no longer just helping us find information. We use it to brainstorm, diagnose problems, and choose between competing options. Increasingly, AI is showing up precisely where there may not be one objectively correct answer.

And this is where the deeper issue begins. 

AI is extraordinarily good at compression. It can take volumes of research, absorb competing arguments, ingest historical context, and give us something coherent enough to understand in minutes.

That is one of the reasons these systems are so useful. But compression always involves a choice about what survives.

When I ask an AI system to explain a difficult question, I usually receive the answer without seeing much of the intellectual landscape that existed on the way to it. I may get a caveat or be told that experts disagree, but I rarely experience the disagreement itself.

Before AI, that disagreement was often hard to avoid.

You read one article and then another and had to decide what to do with the differences.

The contradiction was inefficient, but it also told you where certainty ended. It exposed assumptions and made it obvious that intelligent people could examine similar evidence and still arrive somewhere else.

Looking back, I think we underestimated how much information was contained in that experience.

Disagreement contained information too.

The answer was one part of what we needed to understand. The range of possible answers was another.

That distinction becomes important when several AI systems appear to agree. If five people independently consider a difficult question and all reach roughly the same conclusion, their agreement carries weight because their independence tells us something.

Now ask five AI systems.

They may use different language, emphasize different details and even appear to approach the question differently. But if they are repeatedly drawn toward similar areas of the intellectual landscape, then five answers may not represent five genuinely independent perspectives.

What looks like consensus may sometimes be closer to compression.

The conclusion may still be correct. What becomes harder to see is the intellectual landscape that sat behind it. And consensus is a powerful trust signal.

We have spent our lives learning that independent agreement increases confidence. If several intelligent observers reach the same conclusion, we assume there must be something to it.

But the word doing the most work there is independent.

AI systems do not need to be secretly connected to some universal machine brain for their answers to converge. They may share overlapping training material, similar optimization pressures, common human preferences and other forces that pull their answers toward familiar regions of thought.

The researchers themselves are careful not to claim they know exactly why this happens.

But the consequence is worth considering.

For most of our history, learning more usually meant encountering more disagreement. A library exposed you to competing authors. A field of study exposed you to rival schools of thought. The deeper you went, the harder it became to pretend the answer was simple.

AI creates a strange possibility. We may gain access to more knowledge than any generation in history while encountering less of the disagreement contained inside it. That is a very different information environment.

We have spent centuries worrying about who controls access to information.

We may spend the next century worrying about who controls the range of answers.

There is an irony in what happened next. Researchers are already exploring ways to make AI outputs more diverse, experimenting with methods designed to push models toward different regions of the possible answer space rather than repeatedly returning to the same familiar territory.

After decades of building technology that removes friction from our lives, we may now be discovering that some forms of friction were carrying something we did not know how to measure

The shortcut to efficiency may have inadvertently stunted our journey toward understanding.

Most of the time, making things faster and easier has been the right instinct. But are we   confusing friction with waste?

Some friction is genuinely pointless. But some of it forces us to sit with uncertainty longer than we would like. Disagreement often does exactly that.

None of this means disagreement is inherently valuable. Human beings are perfectly capable of arguing for ideological, performative or simply wrong reasons, and AI should not manufacture contrarian answers just to appear diverse.

The more important question is whether genuine difference remains visible when it exists.

If the evidence overwhelmingly supports one conclusion, say so. But when reasonable interpretations differ, intelligence should preserve more of that landscape rather than collapse it too quickly into one polished answer.

I increasingly think about this when an AI answer arrives a little too easily. I want to know what assumption sits underneath it. What evidence would weaken the conclusion. Whether another intelligent observer could look at the same problem and reasonably see something different.

Not because I want more argument but because I want to know where the edges of the answer are.

Human knowledge has never developed as a clean march toward consensus. It has branched, doubled back and wandered down paths that went nowhere. Schools of thought appeared and disappeared. Ideas considered ridiculous in one generation became foundational in another.

From the perspective of efficiency, much of that probably looked wasteful.

Yet somewhere inside that mess, we kept discovering things the accepted answer had missed.

Perhaps disagreement was never a failure of collective intelligence.

Perhaps it was one of the ways collective intelligence preserved possibility.

And that is what makes the AI hivemind research feel more consequential to me than the phrase itself suggests.

The real concern is not that machines will all begin saying the same thing.

It is that they may become very good at showing us the center of what humanity knows while gradually making the edges harder to see.

Those edges are where doubt survives, where minority opinions persist and where strange ideas often begin. They are also where someone looks at something everyone else accepts and sees it differently.

We have become extraordinarily good at compressing the world’s knowledge.

We should be careful that, in compressing it, we do not also compress the range of human thought.