0
0
Here is a question you almost certainly can’t answer: what colour is the team’s shirt in the film Bend It Like Beckham?
If you’re like most people, the honest answer is “I don’t know.” And that’s a perfectly good answer. It’s true. It costs you nothing. It keeps you from being wrong.
Now picture the same question with one small change. There’s an AI assistant sitting right there on the screen, and it has already offered an answer. Short. Smooth. Sure of itself.
Do you still say “I don’t know”?
A new set of experiments suggests you almost certainly won’t. And the reason why says a lot about how our minds decide what we know.
What the Study Found About Saying “I Don’t Know”
Three researchers in Italy and France — Chiara Marcoccia, Walter Quattrociocchi and Valerio Capraro — ran five experiments with 3,132 people. Four were preregistered, meaning the researchers wrote down what they would test before they saw any results. One was a direct repeat of an earlier experiment, to check the result held up.
The setup was simple. People answered hard questions. They could always skip a question by saying “I don’t know.” Some people had an AI assistant available. Some didn’t.
Here is what happened in one of the experiments, as reported when the paper came out in July:
- Without AI, 44% of answers were “I don’t know.”
- With AI, that fell to 3%.
- Getting it right fell too — from about 27% of answers correct to about 9%.
- Confidence went the other way. On a scale of 0 to 100, it climbed from about 30 to about 76.
Read that again. People got worse at the task. And they felt much better about it.
In the researchers’ own summary, people with AI access “were correct about a third as often” — “yet their confidence nearly doubled.”
The Twist: The AI Was Wrong on Purpose
This is the part most headlines skip, and it changes how you should read the whole thing.
The AI in these experiments wasn’t wrong because AI is always wrong. It was wrong by design.
The researchers chose questions about tiny visual details in films — the colour of a team’s shirt, the car a character drives. These are things AI models are often bad at. They mainly used a model called Step 3.5 Flash, picked because it usually got these questions wrong. They also checked leading models from the big AI companies.
Why set it up this way? Because they wanted to separate two things that normally blur together: using AI and AI being right. If the AI had been correct, you couldn’t tell whether people trusted it because it was good, or simply because it was there.
So the study is not really saying “AI gives bad answers.” It is asking something sharper: What happens to your own judgment when any answer at all is sitting in front of you?
And the answer is: your willingness to say “I’m not sure” nearly disappears.
You Didn’t Even Have to Ask
One more detail from the paper is worth sitting with.
In some experiments, people had to click to ask the AI. In another, the AI’s answer simply appeared on its own, without being asked for. According to a detailed write-up of the results, “I don’t know” answers dropped from 35% without AI to just 1% when the AI’s answer was shown automatically.
Think about how many answers now show up that way. At the top of a search page. In a pop-up. In a suggestion box you never opened. The study’s authors put it plainly: as AI suggestions become everywhere and unasked-for, they may change “the metacognitive threshold at which people decide whether they know enough to answer.”
That’s a big phrase for a simple idea. Everyone has a little inner line. Above it, you feel you know enough to answer. Below it, you say “I don’t know.” A confident answer on the screen seems to quietly move that line.
Did Money Fix It?
The researchers tried the obvious fix. They paid people for right answers and took money away for wrong ones.
It helped a little. People asked the AI less often and got more questions right. In one version, with AI and money on the line, “I don’t know” answers rose to 8% and correct answers to 16%.
But that’s still far below the people who had no AI at all. Even when being wrong cost them, most people found it hard to pass on an answer that was already written out for them.
Valerio Capraro, one of the researchers, told The Register: “For humans, the capacity to say, ‘I don’t know,’ is very important because it represents the recognition of the limits of our own knowledge.” He said he is most worried about children, who are growing up with these tools before they have learned to question them.
What This Study Does Not Prove
It’s worth being as careful with this study as the study asks us to be with AI.
- It is a preprint. It was posted in July 2026 and has not yet been through full peer review.
- The questions were narrow. Film trivia is not medical advice, money decisions or homework. People may behave differently when the stakes feel real.
- The AI was chosen to be wrong. That was the point of the design, but it means the accuracy numbers tell you about this setup, not about how often AI is right in everyday life.
What the study does show, fairly strongly and more than once, is the pattern: when an answer is handy, people stop saying “I don’t know” — and feel more sure of themselves while doing it.
If you want the flip side of this, it’s worth reading about the Dunning-Kruger effect, where the question is how well people judge their own skill — and why even that famous idea is less tidy than its reputation.
How to Keep “I Don’t Know” Alive When You Use AI
None of this means you should stop using AI. It means it helps to protect one small, honest habit. These cost nothing and take seconds:
- Guess first, then look. Before you read the AI’s answer, decide what you think — or decide you don’t know. Now you have something to compare it against.
- Ask “How would I know if this is wrong?” If you can’t answer that, you don’t know yet. You’ve just been told something.
- Treat an AI answer as a lead, not a verdict. A lead is something you check. A verdict is something you repeat.
- Say “I’m not sure” out loud sometimes. At work, at home, with friends. It gets easier with practice, and it gives other people permission to say it too.
- If you have children around you, let them hear you say it. They learn it from watching, not from being told.
It also helps to notice the moments you most want an answer — often they’re the moments you most want to be certain, not the moments you most need to be right. Confirmation bias works on exactly that hunger.
Why “I Don’t Know” Is Worth Protecting
There’s something quietly brave about the words “I don’t know.” They admit a limit. And most of us would rather not.
That’s why a smooth answer on a screen is so tempting. It lets us skip the small discomfort of being someone who doesn’t have it all figured out.
The oldest wisdom people have ever written down keeps coming back to the same odd idea: the person most sure of themselves is often the one in the most danger, and real wisdom starts when you admit where your knowing runs out. The Bible’s old wisdom writers said that, again and again. And there’s a kind of freedom tucked inside it. You were never meant to be the one who knows everything. God already does — which means your “I don’t know” isn’t a failure. It’s just the truth about being human, said out loud.
Maybe that’s the skill worth keeping sharp in the age of instant answers. Not knowing more. Knowing where your knowing ends — and being at peace there.
So the next time a question comes up and an answer is sitting right there, take one breath first. Ask yourself what you actually know. And if the honest answer is “I don’t know,” say it.
It might be the smartest thing you say all day.
Discussion Question
Do you think AI answers are making people more honest about what they know, or less — and have you caught yourself trusting an answer just because it was already there? Tell us in the comments.
Share This
Post 1: A study found that when an AI answer was on screen, people almost stopped saying “I don’t know” — from 44% to 3% — even though the AI was set up to be wrong. Their confidence more than doubled. That’s worth thinking about. https://bgodinspired.com/?p=117022
Post 2: Guess first, then look. That’s my new rule with AI after reading this. https://bgodinspired.com/?p=117022
Post 3: “I don’t know” might be the most underrated sentence of our time. This piece on a new AI study made me think about why it’s so hard to say. https://bgodinspired.com/?p=117022
Questions People Ask About AI and Saying “I Don’t Know”
Does using AI make people less willing to say “I don’t know”?
A 2026 preprint by Chiara Marcoccia, Walter Quattrociocchi and Valerio Capraro, with 3,132 participants across five experiments, found that having an AI assistant available nearly eliminated people’s willingness to say “I don’t know.” In one experiment, “I don’t know” answers fell from 44% without AI to 3% with AI. The effect held whether people asked the AI for help or the AI’s answer was simply shown to them.
Why were people more confident but less accurate with AI in the study?
In the study, the researchers deliberately chose film-trivia questions that the AI usually got wrong, so they could separate the effect of using AI from the effect of AI being correct. People with AI access answered more questions and got about a third as many right as people without AI, yet their confidence nearly doubled — rising from about 30 to about 76 on a 0-100 scale in one experiment. A fluent, ready-made answer appears to lower the point at which people feel they know enough to answer.
Did paying people for correct answers fix the problem?
Only partly. When the researchers rewarded correct answers and penalized wrong ones, participants asked the AI less often, answered more accurately and said “I don’t know” more often. But even with money at stake, people with AI access still declined to answer far less often than people who had no AI at all.
How can I avoid over-trusting AI answers?
A simple habit is to decide what you think before reading an AI’s answer, so you have something to compare it against. It also helps to ask “How would I know if this is wrong?” and to treat an AI answer as a lead to check rather than a final verdict. Saying “I’m not sure” out loud when it’s true keeps the habit of honest uncertainty strong.
Is the AI “I don’t know” study peer reviewed?
As of September 2026, the study “AI advice suppresses people’s willingness to say ‘I don’t know’, even when the advice is wrong and accuracy is incentivized” is a preprint posted in July 2026 and had not yet completed peer review. Four of its five experiments were preregistered and one was a direct replication, which strengthens the core finding, but the questions were narrow film-trivia items and the AI was chosen to be wrong by design.
If this made you think, you might also like: Am I a Bad Person for Using AI? The Question Underneath.
