AI makes us stop saying “I don’t know”, even when it is wrong

1. The experiment in one minute

Researchers from the École Normale Supérieure (Paris), Sapienza University of Rome and the University of Milan-Bicocca ran five online experiments with 3,132 participants in total. Four were preregistered, one was a direct replication.

Everyone answered the same six tricky questions about small visual details in films, for example the colour of the team’s uniform in Bend It Like Beckham or Agatha’s hairstyle in The Grand Budapest Hotel. Crucially, participants could always answer “I don’t know”.

The trick: the questions were chosen so that the AI used in the study (a language model called Step 3.5 Flash) answered them incorrectly almost every time. The authors also tested GPT-5.5, Claude Sonnet 4.6 and Gemini 3.5 Flash: all of them failed on the hardest question. Because the AI was usually wrong, following it could not be explained as a sensible use of a reliable tool.

2. Finding 1: “I don’t know” almost disappears

Without AI, a substantial share of answers were an honest “I don’t know”. As soon as AI advice was available, that share dropped to almost zero in every single study.

Without AIWith AI

Study 1aAI on request (live) · N = 314

36%

6%

Study 1bAI on request (replication) · N = 310

44%

3%

Study 2AI on request · N = 812

17%

1%

Study 3AI on request · N = 843

32%

2%

Study 4AI shown automatically · N = 853

35%

1%

0%10%20%30%40%50%

Show data as a table

Study	Participants	Without AI	With AI
Study 1a (AI on request (live))	314	36%	6%
Study 1b (AI on request (replication))	310	44%	3%
Study 2 (AI on request)	812	17%	1%
Study 3 (AI on request)	843	32%	2%
Study 4 (AI shown automatically)	853	35%	1%

Source: Marcoccia, Quattrociocchi & Capraro (2026), arXiv:2607.13562, Studies 1a–4.

Particularly relevant for everyday life is Study 4: there, participants did not even have to ask the AI. The suggestion simply appeared next to each question, the way AI summaries now appear in search engines or writing tools. The effect was just as strong: 35% “I don’t know” without AI, 1% with it.

3. Finding 2: More confident, less correct

Giving more answers would be fine if they were right. They were not. Without money at stake, participants with AI were correct only about a third as often as those without AI. At the same time, their confidence in their own answers more than doubled.

Without AIWith AI

Show data as a table

	Without AI	With AI
Correct answers (pooled across studies)	27.5%	9.2%
Average confidence, scale 0–100 (Study 2)	29.6	75.9

Source: Marcoccia et al. (2026), Discussion and Study 2. Two separate charts with their own scales.

The authors point out an uncomfortable detail: the AI group answered more questions but got fewer right. So the AI did not just fill gaps. It must have turned some answers that people knew, or could have worked out, into mistakes. Normally, people are rather sceptical of other people’s advice; here they did the opposite and followed a confident but wrong suggestion.

4. Finding 3: Money at stake helps a little

In Studies 2 to 4, some participants earned $0.10 for every correct answer and lost $0.10 for every wrong one; “I don’t know” cost nothing. This changed behaviour, but only partly:

In other words: when mistakes have a price, people use their own judgement more. But even then, “I don’t know” stayed far below the level of people without AI: with money at stake, 41% of answers were “I don’t know” without AI, but only 8% with AI. The authors conclude that AI changes people’s willingness to judge for themselves, not their ability, and that the pull toward relying on AI is strong enough to survive small financial incentives.

5. Why this matters for learning and curiosity

Our view: The study did not measure curiosity or long-term learning. What follows is our interpretation for students, based on what the study did measure.

“I don’t know” is not a failure. It is the starting point of every real question: it is the moment you notice a gap and decide to look into it, check a source, ask a lecturer or think it through. Curiosity begins exactly there.

The study suggests that a ready-made AI answer can skip this moment. The answer arrives before the question has properly formed, and it feels convincing, so the gap seems closed. If that becomes a habit, you lose two things: the chance to discover what you actually don’t know, and the experience of working it out yourself. The authors put it this way: the willingness to say “I don’t know” “may be among the first casualties of human–AI interaction”.

The good news from the same study: the effect is not a fixed limitation. When people had a reason to be careful, they relied on the AI less and got more answers right. Using your own judgement is a choice you can train.

6. Six habits for students who use AI

These are our recommendations, inspired by the findings. They are not results of the study itself.

1. Think first, then ask. Write down your own answer, or “I don’t know”, before you open the AI.
2. Rate your confidence. Note how sure you are (e.g. 30%). If the AI makes you feel 90% sure, ask yourself why.
3. Treat fluency as a warning sign, not as proof. AI answers always sound confident, even when they are invented.
4. Check the source, not the sentence. Ask for references and open them. Invented sources are a classic AI error.
5. Make “I don’t know” part of your work. In a seminar paper, naming an open question honestly is better than covering it with a plausible AI answer.
6. Disclose how you used AI. At many universities, and according to several German administrative courts, undisclosed AI use in graded work can count as cheating. Transparency protects you.

7. FAQ