• Willingness to say 'I don't know' collapsed from roughly 44 percent without AI access to about 3 percent with it, per The Decoder's report on the research.
• Participants using AI reported sharply higher confidence while getting correct answers only about a third as often as those without AI access.
• The pattern held across a five-experiment, 3,000-plus-participant design, including versions with financial stakes and automatic (non-optional) AI answers.
• The underlying research is a PsyArXiv preprint, not yet peer-reviewed, so specific percentages should be treated as reported, not settled.
• For enterprises deploying AI copilots, the finding describes a concrete mechanism — automation bias suppressing hedging — that undermines any workflow assuming a human will flag a wrong AI output.

A study using five separate experiments and more than 3,000 participants found that simply having access to an AI-generated answer was enough to make people almost entirely stop hedging. Willingness to say 'I don't know' fell from roughly 44 percent in the no-AI condition to about 3 percent once an AI answer was available — even though, according to The Decoder's report on the research, the AI was wrong most of the time. The effect held regardless of whether taking the AI's answer actually paid off.

The design: five experiments, one manipulation

The reported design varied whether participants could see an AI-generated answer before responding, while holding the underlying questions constant across conditions. Some versions let participants opt into consulting the AI; others attached small financial stakes to correct answers, and at least one version displayed the AI's answer automatically rather than on request, per The Decoder. Across these variations, the pattern held: the presence of an AI answer, not its correctness, was the variable that predicted whether someone still admitted not knowing.

Confidence rose as accuracy fell

The study reported a widening gap between how sure participants felt and how often they were actually right. Participants with AI access rated their own confidence well above those without it, even as their share of correct answers dropped to roughly a third of the no-AI group's accuracy, according to The Decoder's account. The visible presence of an AI answer inflated self-assessed certainty while degrading the outcome that certainty is supposed to track.

Why this matters for deployed AI systems

The result matters beyond the lab because most enterprise AI deployments rely on a human to catch a wrong model output before it causes damage — an editor overriding a drafting tool, an analyst rejecting a flagged transaction, a support agent correcting a chatbot suggestion. If the mere presence of an AI-generated answer suppresses a human's willingness to flag uncertainty, that human backstop is weaker than governance documentation typically assumes, independent of how accurate the underlying model actually is.

The verdict

This is a Signal, not Noise, for anyone designing human-in-the-loop controls: automation bias can erase the exact behavior — admitting uncertainty — that review processes depend on, and it does so even when the AI is demonstrably unreliable. The finding is preprint-stage and not yet peer-reviewed, so the precise percentages should be treated as reported rather than settled. But the mechanism — confidence rising as accuracy falls once an AI answer is visible — is specific enough to test directly inside any workflow that assumes a human reviewer will catch the model's mistakes.