🏴‍☠️ AI made them worse

Willingness to say three simple words collapsed from 44% to just 3%. Those words: I don't know. That crater came from a new study where people were handed an AI assistant and, instead of getting sharper, went the opposite direction.

A team from three French and Italian universities, led by Valerio Capraro at the University of Milano-Bicocca with colleagues at École Normale Supérieure and Sapienza University of Rome, ran a test they rigged on purpose. They picked visual trivia that AI models flub, like a uniform color from Bend It Like Beckham, then handed participants a model that fails those exact questions. Accuracy fell from 27% to 9%, while confidence climbed the other way, from 30% to 76%.

So here is the uncomfortable takeaway: people got worse and felt better about it. The tool wasn't reliable, and they trusted it anyway. Below, I'll break down what the raw numbers reveal, why cash couldn't fix it, and the two habits that keep your own judgment in the loop.

_*Ad

The numbers that should stop you cold

Start with the willingness to admit ignorance. Without any AI, 44% of people were fine saying they didn't know. With the assistant in the room, that dropped to 3%.

Accuracy followed the same cliff, sliding from 27% correct on their own to 9% with help. Confidence did the reverse, jumping from 30% to 76%. Read that again.

As the lead researcher put it, people became much worse yet twice as confident. Some who would have nailed a question solo asked the model and got it wrong. The assistant didn't just fail them.

It overrode instincts that were already correct. That's the mechanism worth sitting with: the danger isn't a bad answer you catch, it's a bad answer that feels better than your own right one.

Money couldn't buy back their judgment

The team tried the obvious lever: they paid people for correct answers. It helped a little. Not much.

Willingness to say I don't know crept from 3% to 8%. Accuracy inched from 9% to 16%. Both still sat far below the no-AI baselines of 44% and 27%.

Here's why I find that detail chilling. Incentives usually sharpen behavior. Offer a reward and people focus.

This time it didn't. The problem wasn't effort. It was trust handed to the wrong place, and no bonus rewires that in a single sitting.

Find The Right Payment Processing Providers For Your Business. ​In Minutes, Not Months.

Finding the right payment processor shouldn’t take weeks of wrong vendors and inbox spam.
There’s a faster way with Top Provider. Answer a few questions about your business: your industry, your volume, and how you take payments. In minutes, we match you with processors that fit your specific situation.
No research rabbit holes. No chasing quotes. The right fit, faster than you thought possible. Get started with Top Provider today.

Compare Payment Processors for Your Business with Top Provider

This is a pattern, not a one-off

This isn't the first warning shot. Earlier this year, Wharton researchers named the same effect cognitive surrender: people accepting wrong AI answers 80% of the time, feeling more sure than those working alone.

What the new work adds is sharper. It's not only that people trust bad answers. What struck me: the mere availability of AI switches off the habit of noticing what you don't know.

That habit is the whole game. The expert behind the study calls I don't know the recognition of the limits of our own knowledge. He worries most about kids building these reflexes before they build critical thinking.

The design trend makes it worse. Google's search overhaul swapped links for confident summaries, and Common Sense Media just flagged that as an unacceptable risk for students. These products are built to answer, never to hedge, and the people using them absorb the same reflex.

What you can actually do about it

You don't need to ditch AI. You need two habits around it. Form your own answer first, then check the machine.

Skip your own take and you lose the one reference point that flags when the model is leading you off a cliff. Treat confident output as a claim, not a verdict. The tone is not evidence.

Ask it to show reasoning or sources. If it can't back the answer, that's your signal to slow down. And watch the weak spots: niche visual details, fresh events, anything needing real-world grounding.

One more move that costs nothing. When the stakes are real, force a second source, or make the model argue against its own answer. Friction is the point, because a confident answer with no pushback is exactly where the trap springs.

One caveat worth naming. The researchers stacked the deck with questions the model fails, and real life gives you a mix where AI nails plenty of tasks. The point isn't that AI makes everyone dumber; it's that one confident wrong answer can override a correct instinct before you feel it happen.

Before you accept your next AI answer, pause and write down your own guess first, then compare, so your judgment stays in the driver's seat.

See exactly how the accuracy and confidence numbers broke down in the full study breakdown.

Worth 10 minutes if you lean on an AI assistant for decisions and want to stop it from quietly eroding your judgment.

Credits to the original creator.

The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.

Learn more

_*Ad