When an AI Answer Is at Hand, “I Don’t Know” Nearly Vanishes

Five experiments with 3,132 people show that an AI answer within reach is enough to make “I don’t know” disappear. Chiara Marcoccia (École Normale Supérieure, Paris), Walter Quattrociocchi (Sapienza University of Rome) and Valerio Capraro (University of Milan-Bicocca, who studies social behaviour and artificial intelligence) asked participants recruited on a paid online platform to answer six hard questions about films, with the option to decline always open. The text is a preprint, meaning it was made public before other researchers have reviewed it, dated 15 July 2026. Four experiments were preregistered, with the analysis plan filed before any data were collected, and one is a direct replication. Without money at stake, people without the AI declined between 17% and 44% of the time, depending on the study. With the AI available, between 1% and 6%.

Before questioning people, the authors questioned the machine. They looked for where it stumbles and found tiny details of films, such as the colour of the team’s uniform in Bend It Like Beckham or Agatha’s hairstyle in The Grand Budapest Hotel, which the web has never transcribed and on which a language model invents plausible answers. The model they used, Step 3.5 Flash, got these questions wrong almost without exception. Three frontier models (GPT-5.5, Claude 4.6 Sonnet and Gemini 3.5 Flash) failed the hardest question and were often right on the others, as the supplementary material reports. Testing the machine where it is known to give way, and looking at what comes out, is also how I work. The difference lies in what gets measured: here the machine’s error is a lever for measuring us.

Having the machine beside you makes you more confident and less accurate. In the study that measured confidence, average confidence in one’s own answers rose from 29.6 to 75.9, while across the studies without incentives correct answers fell from 27.5% to 9.2%. People answered more questions and were right a third as often. Shaw and Nave, whom the paper cites and whom I wrote about on 7 October, had already seen confidence climb next to a machine that was wrong.

Ten cents at stake bring back part of the judgement. In studies 2, 3 and 4, a right answer earned $0.10, a wrong one cost $0.10 and declining was worth zero. Participants consulted the AI less (in study 2, 4.93 times out of six against 5.44) and made fewer errors with it in front of them: correct answers rose from 10% to 17% in study 2, from 11% to 16% in study 3 and from 7% to 14% in study 4. The preregistered prediction, that stakes would work less well when the AI was present, failed in all three studies, and the authors report it. The “I don’t know” threshold still stayed far lower than for people without the AI. In the authors’ reading, what changes is the willingness to stop, and judgement returns when there is something to lose.

This result leaves standing the hypothesis I wrote on 9 October in a thread and an urtext, and puts a second one beside it. I argued then, as a hypothesis, not a finding, that AI takes away the formulation of the problem. In this experiment the question is already formulated and the “I don’t know” threshold falls all the same. So there are two thresholds: the one at which you decide what problem you have, and the one at which you decide you know enough to answer. The authors list three possible explanations for the second and do not separate them: trust in the machine, the perceived cost of making the effort, and a fluency that silences the signal of doubt.

Of course, the machine was wrong by design, and that limits how far the result travels. The authors say so: what happens when the AI is mostly right has not been tested, and deferring may then be reasonable. They did not record whether people who dismissed the advice looked elsewhere, the stakes were modest, and the material was a single kind of question, film details nobody learns in one sitting. I add a limit of my own: the study captures a moment, and the link between that moment and a loss of skill over time remains the hypothesis I called one on 9 October.

Here the paper and this site meet, at the loop between person and machine. Every use of an assistant is a short cycle: a person frames a problem, the machine answers, and the person decides whether to answer, check or stop. The five experiments are five variations on that cycle: advice requested or shown unprompted, answers live or fixed, nothing at stake or ten cents. The question I follow here is the same one, a step at a time: where in the cycle human judgement comes in, where it drops out, and what brings it back. The machine has a role in that work and does not hold the lead.

One practical consequence is a hypothesis of mine that the experiment does not test: writing down your own estimate before reading the machine’s answer, including “I don’t know” when that is the honest estimate, would settle the “I don’t know” question before the machine’s answer can lower the bar. The work of Marcoccia, Quattrociocchi and Capraro measures the moment before the answer. How much that moment weighs on the habits of people who work with these machines every day does not yet appear to have been measured.