AI access cuts "I don't know" to 3% even when AI is wrong

The study sits in a specific research lineage the authors invoke: "Epistemia," the tendency to accept AI answers because they sound convincing rather than because they are checked. A language model always has to produce an answer and never pauses when it doesn't know something; the authors suggest users who hand off judgment to such a system may adopt its lack of restraint. That reading is reinforced by how far the results run against the existing "advice use" literature, where people normally underweight outside advice and shift only about a third of the way toward an advisor's position — the opposite of what participants here did.

The five experiments also isolate what does and doesn't move the effect. Financial incentives reduced how often participants sought AI advice and improved correctness when AI was available, but across Studies 2 through 4 none of the pre-registered tests showed a statistically significant interaction between incentives and AI availability — the two levers appear to work largely on their own. Study 4's automatic-display condition, which mirrors search engines that show AI summaries unprompted, barely changed the result, suggesting the effect isn't just about people actively choosing to consult a tool.

The stakes, on the authors' own framing, may lie less in model accuracy than in something harder to engineer: whether people keep recognizing the limits of their own knowledge as AI answers become ubiquitous and increasingly unsolicited. For organizations that rely on employees flagging uncertainty, the open questions the researchers leave — whether the effect holds beyond movie trivia, and whether larger or reputation-based incentives would close the gap — are likely to matter more than any further gain in model accuracy.