Give people access to an AI assistant and they almost stop admitting when they don’t know something. They also get more answers wrong. The Decoder reports on a new study by Marcoccia, Quattrociocchi, and Capraro (2026). In it, participants with AI access became far more confident and far less accurate on movie trivia questions, even though the AI’s advice was mostly wrong.
That last detail matters. People weren’t sensibly handing work to a tool they could trust. They were deferring to a system that kept misleading them.
📊 The numbers
The researchers ran several experiments. Each one compared a control group with no AI to a group that could get AI advice. Participants could answer a question or pick “I don’t know,” which the researchers call suspending judgment.
How often people said “I don’t know”:
In Study 1a, people with no AI access said “I don’t know” 36% of the time, while those with AI access said it only 6% of the time. Study 1b showed even starker numbers: 44% without AI dropped to 3% with it. Study 4, where AI was shown automatically with no incentives, showed 35% without AI falling to just 1% with it. When incentives were added in Study 4, the no-AI group was around 39%, while the AI group rose to 7%.
Study 2 compared confidence against accuracy (no incentives):
Confidence rose from 29.6 to 75.9 out of 100, roughly 2.5 times higher. But correct answers fell from 27.6% to 10.0%. Across all studies without incentives, participants got 27.5% of questions right without AI and only 9.2% with it. They answered more questions but were right about a third as often. The authors conclude that AI access turned some answers that would have been correct into errors.
💰 Money helps, but not much
Studies 2 through 4 added financial stakes. Participants earned 10 cents for a correct answer, lost 10 cents for a wrong one, and got nothing for “I don’t know.”
The researchers had pre-registered a prediction: incentives would make people more willing to abstain, but AI access would weaken that effect. The data didn’t back up the second part. None of the three studies found a statistically significant interaction. The two factors mostly worked on their own: with incentives, people asked the AI a bit less often (4.53 vs. 5.27 out of six possible requests in Study 3), they answered correctly more often when AI was available, and their willingness to say “I don’t know” went up slightly but stayed well below the no-AI group. The researchers call the incentive effects modest. They don’t know whether bigger rewards, or reputation-based ones, would close the gap.
🤖 Unsolicited AI works just as well at this
What stands out most to me is Study 4. Here the AI answer showed up automatically, and nobody had to ask for it. The researchers say this mirrors how AI now reaches people: search engines put AI summaries at the top, and writing tools suggest text you didn’t request. The effect barely changed. Dropping from 35% “I don’t know” to 1% is close to total collapse, and it happened without anyone asking the AI anything.
🧠 Why this happens
The authors tie the results to what they call “Epistemia,” the habit of accepting AI answers because they sound convincing, not because anyone checked them. A language model always produces an answer. It never stops to say it doesn’t know. When people hand their judgment to a system like that, they may pick up its lack of restraint. This also cuts against decades of research on how people take advice. Normally, people underweight outside input and move only about a third of the way toward an advisor’s view. With AI, participants did the opposite.
⚠️ Limitations
- The task was movie trivia, and the researchers say it’s unclear whether the effect holds as strongly in other areas
- The AI advice in the study was mostly wrong, which may not match how people use AI in real life
- The financial incentives were small, so stronger stakes haven’t been tested
What to do with this
If you use AI for research, decisions, or client work, a few habits follow from this study:
- Make abstaining easy. In team workflows, treat “I’m not sure” as a real answer, not a failure.
- Rate your confidence before you check the AI. Notice when an AI answer bumps your certainty without adding any evidence.
- Be careful with unsolicited answers. AI summaries you didn’t ask for still pull your judgment.
- Explain, don’t delegate. Other research cited by The Decoder found that 10 to 15 minutes with an AI assistant can reduce problem-solving ability and persistence on later tasks. People who used the AI only for explanations, or ignored it, showed no decline.
This study fits a growing body of work. A Swiss Business School study of 666 participants found a strong negative link between AI use and critical thinking, and Microsoft research points to heavy demands on metacognition, the ability to monitor your own thinking. The researchers warn that as AI answers become ubiquitous, the willingness to say “I don’t know” could be among the first casualties. Their point is that protecting human judgment may depend less on making models more accurate and more on people still recognizing where their own knowledge ends. The full breakdown is available at The Decoder.