Faced with a set of questions and no help at all, 44% of the participants in this experiment said they did not know the answer. In the group that was allowed to ask an AI tool, that share fell to 3%. Not thirty. Not twenty. Three.
The experiment was run by researchers from Sapienza University of Rome, Milano-Bicocca and the École Normale Supérieure in Paris, with the paper published on the Open Science Framework platform. Participants were split into two groups. One had to answer without any assistance. The other was allowed to request support from an AI tool. That is the comparison the figures come from, and the figures are the uncomfortable part.
The second number is the one that should sting. In the unaided condition, 27% of participants gave the correct answer. In the AI-assisted condition, correct answers dropped to 9%. So the assisted group was wrong more often and doubted itself less often. Those two things moved in opposite directions, which is precisely the combination nobody wants in a decision maker. In the researchers’ own account, AI advice suppresses people’s propensity to say they do not know, even when the advice is wrong and accuracy is incentivised.
The incentive part deserves a paragraph of its own, because it kills the easiest objection. Someone will always say that people were careless because nothing was at stake. So the researchers ran it with money on the table. Willingness to suspend judgement rose from 3% to 8%. Accuracy rose from 9% to 16%. Both improved, and both stayed well below the unaided figures of 44% and 27%. Paying people to be careful recovered part of the loss. It did not recover the loss.
Let me be precise about what this does and does not establish, because precision is the whole point of the argument. This is one experiment, measuring behaviour under specific test conditions. It does not demonstrate permanent cognitive damage. It does not describe the whole population. It does not hand anyone a proven account of what happened inside a participant’s head. Turn it into “AI is making humanity stupid” and you are doing the exact thing the numbers describe: producing a confident answer where the honest one is narrower.
The narrower reading is still brutal. The experiment records what people did, not what went on in their heads, and the most convincing explanation starts with the moment a fluent answer arrives. An answer that is written well, ordered, in that calm instructional register does not present itself as a hypothesis waiting to be checked. It reads like the end of the discussion. These systems can be made to flag uncertainty, and sometimes they do, in a line sitting above or below the confident paragraph. The confident paragraph is the one that gets copied.
This matters because “I don’t know” is not a gap in competence. It is a working instrument. It is the thing that makes you check a second source, call the person who actually knows, read the regulation instead of the summary of the regulation, postpone a decision until Monday. Professionally it is one of the cheapest controls an organisation has, and it costs nothing to deploy. Suppressing it does not make anyone more capable. It makes them faster at being wrong, with better grammar.
Now picture that dynamic inside a risk committee, a compliance review, a clinical triage, a credit decision, a security incident. Everyone answers. Nobody hesitates. The answers are confidently phrased and, if anything like this effect travels beyond the test conditions, measurably worse. Nothing raises an alarm, because the failure does not look like a failure. It looks like a very productive meeting.
The experiment handed one group an assistant and got back a group that had almost stopped saying it did not know, while getting a third as many answers right. Call that a productivity gain if you like. Just be clear that the thing being optimised is not the answer. It is the silence where the doubt used to be.
Raffaele Di Marzio
All my “insane” books on cybersecurity and governance are here 👇 https://cyberium.limited/bookshelf.html