Insights
Why We Stopped Saying I Don't Know
There is a sentence that has quietly stopped appearing in meetings. “I don’t know.” Not the performative version, said before an answer arrives anyway. The real one. The pause where someone admits the question is harder than the room assumed, and the group has to slow down. I have been trying to understand why it disappeared, and I think the answer is more interesting than corporate confidence culture.
In a set of five experiments published this year (N = 3,132, four of them preregistered), researchers gave people difficult questions and always allowed them to decline to answer. The questions were engineered so that the available AI advice was wrong — which separates the use of AI from the accuracy of AI.
Merely having access to an assistant
Merely having access to an assistant nearly eliminated participants’ willingness to suspend judgment. They answered far more questions and were correct about a third as often. Their confidence roughly doubled. Read that again, because the striking finding is not the error rate. It is the confidence. The tool did not only change what people believed. It changed the threshold at which they felt entitled to believe anything at all.
A second study points at the mechanism from another angle. In a competitive question-answering setting, expert humans working with AI agents made better decisions overall than either alone — but their mistakes were systematic. Under-reliance was highest, at roughly 61 percent of opportunities, precisely when the AI’s suggestion agreed with the human’s own initial wrong answer. Agreement was treated as evidence. Confirmation felt like confirmation.
So the machine that was supposed to expand our epistemic range is, in practice, often narrowing it — not by lying, but by being fluent, immediate, and agreeable.
I want to be careful here. This is not an argument that AI makes people stupid. In both studies, the collaboration outperformed the individual. And in the first, when accuracy was rewarded and error penalised, people sought advice less, followed it less, and admitted uncertainty more often.
Incentives moved behaviour
Incentives moved behaviour. That matters enormously, because incentives are the one variable organisations actually control. Which brings me to the part that keeps me up. Most organisations have spent decades penalising uncertainty. Not officially — no policy says so — but in the small economics of meetings: who gets interrupted, whose forecast is remembered, whose “let me think about that” reads as unpreparedness. We built environments in which not knowing was expensive. Then we handed everyone a device that makes not knowing entirely avoidable.
Fluency has become abundant
The result is an organisation where every question has an answer, arriving faster than the deliberation that would have tested it. Fluency has become abundant. What has become scarce is the willingness to stand in front of colleagues and say the sentence that stops the machine.
I have argued before that AI relieves execution and relocates the human constraint. This is the same claim at the level of the individual mind. When the cost of producing an answer falls to nearly zero, the only remaining source of quality is the judgement about whether an answer should be produced at all. That judgement is not a skill. It is a permission.
Which means the leadership task here is not training. It is licensing. I wonder what would change if “I don’t know, and here is what would tell us” were treated as a contribution — logged, valued, referenced later when it proved right. Not as humility theatre, but as the organisation’s most honest early warning system. The person who suspends judgement is doing something no assistant currently does: they are protecting the group from a confident answer that nobody has yet earned.
Psychological safety was always an economic argument. It is becoming an epistemic one.