Market Signals

When confidence stops telling you the truth

Something lands on your desk and you decide whether to trust it. New research suggests the cue you have always used for that has stopped working.

Grace Ge · August 2026

You have a working method for this, whether or not you would put it in words. Someone presents. If they hedge — this number needs another look, I'd want to check that assumption before we commit — you slow down. If they are clear and complete and sure of themselves, you move.

That method works because hedging is expensive. People don't volunteer doubt for fun. When it shows up, it usually means something.

A July 2026 preprint, reporting five experiments with 3,132 people, suggests that both halves of that method are now unreliable.¹

What they did

Participants were given six factual questions and allowed to decline — "I don't know" was always available, and in some conditions it paid better than guessing wrong. Some had an AI assistant. Others didn't.

The questions were picked so the model would get them wrong: obscure film details, the kind of thing that barely exists in web text, so the model invents an answer and sounds certain doing it. That is the important part. Because the advice was usually bad, any change in behaviour can't be waved away as people sensibly trusting a good tool.

And in the final experiment they removed the choice. The AI's answer simply appeared beside every question. Nobody decided to use it.

First: the hedging disappeared

Across the studies, the share of answers where people held back fell from roughly 36–44% to 3–6%.²

The mechanism isn't mysterious. A model always produces an answer. It has no way to sit out a question it can't handle. Hand someone a fluent, complete, confident-sounding response and most of them pass it along — and the "I'm not certain about this one" that would otherwise have travelled with it never gets written down.

The uncertainty didn't go anywhere. Only the disclosure of it did.

Second: confidence stopped meaning anything

This is the finding that should concern anyone who receives work rather than produces it.

People with AI were roughly two and a half times as confident and got about a third as many right.³ That is not a small drift in calibration. In the AI condition, the relationship ran in the opposite direction.

Look at the green line first, because it is the one you have been relying on your whole career. It slopes up. Left to themselves, people are decent judges of their own reliability — in this study their average confidence landed within a few points of their actual accuracy.⁵ Someone who sounds sure usually is.

The other line slopes down. With AI in the room, the people who said they were sure were less often right than the people who said they weren't. Confidence had stopped tracking whether they were correct and started tracking something else: how completely they had accepted what the model told them.

What this is not

It is tempting to read all this as carelessness — people not trying hard enough, in need of a policy about using AI responsibly. Two things in the data argue against that.

Two of the experiments paid for accuracy: ten cents for a right answer, ten cents deducted for a wrong one, nothing either way for declining. Direct, immediate, personal. It helped, and it came nowhere close to restoring the baseline — holding back rose from 1.2% to 7.1% among people with AI, against 34.5% in the group without it.⁶

And in the final experiment nobody chose to consult anything. The answer was simply on the screen. The effect held. Which describes most of how AI now reaches people at work: the search summary above the results, the suggestion inside the document, the reply already drafted in the box. There is no moment at which someone decides to lean on it — so there is no moment at which a policy about leaning on it carefully can bite.

What it is

A measurement problem, and yours to solve rather than theirs.

You have been reading confidence as evidence. That was reasonable, and it was cheap — it cost no time and no process, which is exactly why it became the default. It is now unreliable in a way you cannot detect from the outside, because the fluent unhedged version and the carefully checked version look identical on arrival.

What still works is anything that asks about the process rather than the conclusion. Which part of this would you defend if it turned out to be wrong. What did you check yourself. Where would this break first. Those questions aren't about catching anyone out. They restore information that used to arrive on its own and no longer does.

The uncertainty is still there. It just doesn't announce itself anymore.