Everyone Is Wrong in the Same Direction Now

Something structural changed in the last two years, and I do not think we have priced it correctly yet.

For most of the industrial era, human judgment was diversified by accident. A thousand analysts looked at the same market and produced a thousand slightly different readings, because each of them had a different mentor, a different set of scars, a different bad year that taught them to distrust a different thing. Most of them were wrong. That was fine. They were wrong in uncorrelated directions, and the disagreement did real work — it kept any single error from becoming everyone's error at once.

That accidental diversity is now being removed on purpose, and we are calling it progress.

A very large share of the world's first drafts, opening arguments, initial screens, preliminary diagnoses and strategy decks now pass through a handful of foundation models. They are trained on overlapping corpora. They are tuned toward overlapping notions of what a good answer sounds like. And they are converging, because everyone is chasing the same benchmarks, hiring from the same fifty labs, and reading the same papers on the same weekend.

The risk is not error. It is correlation.

Here is the part that makes this difficult to argue in a meeting: the models are good. On a great many tasks they outperform the median human doing the same work, which is precisely why adoption has been so fast and so uncritical. The case against monoculture cannot be that the crop is bad. It is usually an excellent crop. That is why you planted so much of it.

The case is about what happens on the day something goes wrong.

I spent most of my career in payments and financial infrastructure, so my instinct for this comes from risk, not from computer science. Risk people learned this lesson the expensive way. A portfolio in which every position is hedged the same way is not a hedged portfolio. A banking system in which every institution passes the same stress test has not eliminated stress; it has arranged for every institution to fail on the same input. Diversification is not about the quality of any individual bet. It is about whether your bets can all be wrong simultaneously.

Apply that lens to cognition. If six teams in a company independently research a decision, and all six route their research through the same assistant, the company does not have six opinions. It has one opinion wearing six badges. Worse, it has one opinion that now looks corroborated, because six people arrived at it separately and each of them experienced the process as their own thinking.

Corroboration is the thing that quietly breaks. We use agreement as a proxy for truth because, historically, independent minds agreeing was expensive and therefore meaningful. When agreement becomes free, the signal it carried disappears — but our instinct to trust it does not.

Where this shows up first

I would expect the earliest visible damage in fields where the inputs are ambiguous and the outputs are consequential: credit, hiring, underwriting, security triage, medical differentials, and the whole machinery of corporate strategy.

Think about what a shared blind spot does to a credit market. If the assumption that a certain kind of borrower is safe is embedded in the reasoning of every institution that evaluates borrowers, that assumption stops being an assumption and becomes the shape of the market itself. Nobody is checking it, because everyone's check produces the same result. The error is not detected until it is realized — that is, until money is gone.

Or consider security. Defenders increasingly use models to decide what looks anomalous. Attackers increasingly use the same class of models to decide what looks normal. That is not a fair fight or an unfair one; it is a fight where both sides share a definition, and the exploit lives inside the shared definition, invisible from either chair.

The pattern is the same in each case. When the reasoning layer becomes shared infrastructure, its limitations stop being individual mistakes and become systemic properties. And systemic properties do not announce themselves gradually. They hold, and hold, and hold, and then they do not.

What I would actually do

I am not arguing that anyone should use these tools less. I use them daily, openly, for real work. The argument is about structure, not abstinence.

Track where your judgment is correlated, not just where it is automated. Most organizations have a decent inventory of which tasks are AI-assisted. Almost none have an inventory of which decisions share an upstream reasoning dependency. Those are different maps, and the second one is the one that tells you your exposure.

Buy dissent deliberately, because you will not get it for free. A second model from a second lab is a start, though a weak one — the training data overlaps more than the marketing suggests. Better is a person with standing, time, and explicit permission to say that the entire framing is wrong. That role used to exist informally, distributed across everyone who had to think from scratch. It now has to be someone's actual job.

Keep at least one process where a human reaches a conclusion first and consults the machine second. Not out of nostalgia, and not because the human will be more accurate — often they will not be. Because the order of operations determines whether you get a second sample or an echo. Once you have read the model's answer, you cannot un-read it, and whatever you produce afterward is a variation on it.

Write down what you believe before you ask. This is the individual version of the same discipline, and it costs about ninety seconds. It is the only reliable way I have found to notice, later, that my opinion changed without my ever deciding to change it.

The failure will look responsible

The version of AI risk that gets attention is dramatic: a system with goals of its own, an intelligence that slips its leash. I understand the appeal of that story. It has villains.

The failure I actually expect is drearier and much more likely. It looks like a great many careful, competent, well-governed organizations making the same reasonable-sounding mistake in the same quarter, each of them able to produce impeccable documentation showing that they followed a rigorous process. No villain. No breach. Just a civilization that gradually replaced a noisy, redundant, wasteful diversity of judgment with something faster and cleaner, and did not notice that the noise had been doing something.

Redundancy always looks like waste right up until the moment it is the only thing keeping you alive. We are in the part of the story where it still looks like waste.