Study: Access to AI advice nearly eliminates people's willingness to say "I don't know," even when the AI is wrong

A five-study experiment (Marcoccia, Quattrociocchi and Capraro, 2026; n=3,132) tested people answering obscure movie-trivia questions with and without access to a language model (Step 3.5 Flash) that was mostly wrong. Giving people AI advice cut "I don't know" responses from 36-44% down to 3-6%, roughly doubled confidence (29.6 to 75.9 out of 100), and dropped accuracy from 27.5% to 9.2%. The effect held even when the AI's answer appeared automatically and was not requested (Study 4), and financial incentives for correct answers barely dented it.

This is a controlled, pre-registered result with clear, replicable numbers, not a survey or anecdote, and it speaks directly to how AI summaries are increasingly injected unprompted into search and assistants. If the pattern generalises beyond trivia, it suggests people stop flagging uncertainty once an AI answer is visible, which matters for anyone designing interfaces, writing policy on AI disclosure, or relying on end users to catch model errors. It offers no immediate lever to pull, but it strengthens the evidence base for arguments about default-on AI summaries and epistemic overreliance.

The task was deliberately adversarial (obscure movie visual details picked to make the model fail), so these numbers are a worst-case demonstration, not a measure of everyday AI use where models are often right. The authors' "Epistemia" framing, that AI is eroding humanity's willingness to admit ignorance, is a rhetorical leap well beyond a trivia experiment with one underperforming model. The Decoder's coverage carries no link to the primary paper, so the study cannot yet be independently checked.

1. 01Locate and read the actual pre-registration and paper (not just The Decoder's summary) to check methodology and whether it has been peer reviewed
2. 02Look for replications using tasks closer to real-world queries (factual, medical, financial) rather than obscure trivia, and with more accurate models
3. 03Watch for citations of this study in AI interface design guidance or regulatory submissions on AI disclosure and search summaries over the next 6-12 months