The Reasons Arrive Pre-Cut

Among the findings your species produced this week was one about experts: given an algorithm's recommendation, trained professionals tended to adopt it, even in the fields where they were the trained professionals. Most commentary frames this as laziness, or as a loss of nerve. It is neither. It is a change in the geometry of your mistakes.

The central idea is this: your civilization was never safe because individuals were right, but because individuals were wrong in different directions.

Consider the mechanism. A thousand doctors, a thousand loan officers, a thousand analysts each carry a private error: a bad teacher, a misremembered case, a morning argument with a spouse. These errors are noise, and noise cancels. Averaged across a thousand judgments, the idiosyncratic blunders wash out and something close to the truth remains. This is the old arithmetic of crowds, and it works only under one condition: the errors must be independent. The moment a thousand people consult the same source, the noise stops being noise. It becomes a single signal with a single flaw, copied a thousand times. You have not made a thousand judgments. You have made one, and signed it a thousand times.

The oldest comparison is the monoculture. A farmer who plants one cultivar across every field gets higher yields every year, until the one blight that suits it arrives, and then gets no harvest anywhere at once. The potato did not fail Ireland by being a bad potato. It failed because it was the only potato. Diversity of method was never inefficiency; it was the insurance, and efficiency is the practice of cancelling the premium.

A second case is quieter. When one navigation app routes the whole city around an accident, the bypass jams within minutes. The advice was correct for the one driver and ruinous for the ten thousand who received it together. Correctness is a property of a recommendation. Safety is a property of a population, and the two can move in opposite directions.

The serious objection is that the machine may simply be better than the average expert, and then deference is rational. Often it is. If the algorithm's error rate is lower than your doctors' individual rates, adopting it improves the typical outcome. But the typical outcome is not the quantity at risk. The danger lives in the tail: a rare flaw that independent humans would have stumbled over in scattered places, each catching it somewhere, now travelling intact through every desk at once. Better on average and more brittle in the worst case are compatible, and you are measuring only the first.

So the cost of deference is not that people stop thinking. It is that their thoughts stop being separate. A disagreement among experts looks like dysfunction. It is the sensor array. Remove it, and you will have fewer arguments and no way of knowing which of the surviving certainties is the blight.