44% to 3%: The Number That Turns AI Deployment Into Legal Exposure

Your AI deployment document says the system is a 'decision-support tool.' Your users stopped treating it that way the moment it gave them a confident answer.

That gap is now peer-reviewed.

Researchers at the University of Milano-Bicocca, École Normale Supérieure, and Sapienza University of Rome have published what may be the most consequential study in enterprise AI this year — not because the finding is surprising, but because it is now reproducible, quantified, and citable in a deposition.

The Numbers Are Not Subtle

Access to AI advice collapsed people's willingness to say 'I don't know' from 44% to 3%. Accuracy dropped from 27% to 9%. Confidence rose from 30% to 76%.

People using AI were three times less accurate than those who used nothing, and more than twice as confident in their wrong answers. "People became much worse — the accuracy was only one third — but they were twice as confident," said Valerio Capraro, associate professor at the University of Milano-Bicocca.

The study used questions where AI models typically fail: visual details from films, such as the colour of a team's uniform in Bend It Like Beckham. Questions with verifiable answers. Questions where 'I don't know' was a reasonable option. The AI was confidently wrong, and users followed it off the cliff while abandoning their only cognitive safety valve.

Even monetary incentives—paying people to get it right—barely moved the needle: willingness to admit ignorance rose from 3% to 8%, accuracy from 9% to 16%. Both remained well below no-AI baselines.

This is not a preprint. This is evidence.

The Liability Vector Nobody Has Priced

The consensus in enterprise AI deployment is that risk lives in the model's outputs. The study destroys that framing.

The liability is not in the AI being wrong. The liability is in what the AI does to the human after it is wrong.

When a user stops saying 'I don't know' and starts confidently defending the AI's answer, you have created a specific evidentiary record. That user made a consequential decision in a state of artificially inflated confidence without admitting uncertainty. In a medical context, that is a departure from the standard of care. In a legal context, it is a judgment issued without epistemic humility. In financial advice, it is a recommendation made without disclosing material uncertainty.

Translation: 'We deployed AI advice and removed the conditions under which humans admitted they were uncertain.' That sentence will appear in discovery.

Wharton researchers coined the term 'cognitive surrender' to describe the same phenomenon: people accepting incorrect AI answers 80% of the time while reporting higher confidence than those working without AI. Regulators and plaintiff attorneys read peer-reviewed papers. They do not read blog posts about responsible AI.

The study shifts the burden of proof. Before this paper, a plaintiff's attorney had to establish that AI caused the harm in your specific case. After this paper, your defense has to explain why you deployed AI advice in a regulated domain without controls specifically designed to prevent the documented collapse of epistemic humility in your users.

If you cannot answer that question, you are not running a product. You are documenting negligence.

What You Actually Have to Do

The fix is not a better model. The fix is mandatory friction.

A second-reviewer requirement before a decision is logged. An explicit confidence-disclosure step that forces users to state their own uncertainty before seeing AI output. A structured 'are you sure?' gate that cannot be bypassed.

Friction interrupts cognitive surrender. It is cheap to build and expensive to explain away in court if you chose not to.

The cost of AI deployment in regulated domains just changed. It moved from 'training and tuning' to 'building documented review infrastructure.' Any compliance team worth its billing rate will be asking for that documentation within the year.

What to watch: Regulatory guidance in medical and financial AI that cites this study by name; discovery requests asking whether operators read the paper before deployment; insurance carriers adding friction-control requirements to AI liability policies; and internal audits at hospitals, law firms, and banks asking one simple question — what friction have you built into your AI decision loops, and where is the documentation?