研究发现：有 AI 可用时，人们几乎完全不愿说“我不知道”

Researchers ran five experiments with 3,132 participants to test whether access to a language model changes how people handle uncertainty. Just having AI advice available nearly wiped out people's willingness to say "I don't know." The researchers picked questions where the model they used, Step 3.5 Flash , was almost always wrong. These were questions about fine visual details from movies, like the color of a team uniform in Bend It Like Beckham. According to the authors, such details rarely appear in online text, which makes them prime targets for hallucinations. GPT-5.5, Claude 4.6 Sonnet, and Gemini 3.5 Flash got most of the other questions right but sometimes failed on the harder ones. Because the AI advice was mostly wrong, the effect can't be explained as reasonable delegation to a reliable tool. The researchers still see a tendency for people to defer to AI answers, but whether this holds just as strongly beyond movie trivia remains an open question. In the first two studies (1a and 1b), participants could choose whether to ask an AI for advice. In the control group without AI access, they withheld their judgment on 36 and 44 percent of questions. With AI access, those numbers dropped to 6 and 3 percent. In Study 2, the researchers also measured participant confidence. Without financial incentives, confidence with AI access hit 75.9 points on a scale of 100, compared to 29.6 points without AI. That's roughly two and a half times higher. At the same time, the share of correct answers fell from 27.6 to 10.0 percent. Participants were far more confident but wrong much more often. Across all studies, participants without incentives who had AI access got 9.2 percent of questions right, versus 27.5 percent without AI. They answered more questions with AI but were correct only about a third as often. Studies 2 through 4 added financial incentives. Participants earned 10 cents for each correct answer, lost 10 cents for each wrong one, and got nothing for "I don't know." The researchers had pre-registered the hypothesis that incentives would increase willingness to abstain but that AI availability would weaken that effect. None of the three studies showed a statistically significant interaction. AI access and financial incentives worked largely on their own, according to the study. Incentives led participants to seek AI advice less often, with 4.53 versus 5.27 out of six possible requests in Study 3, and to answer correctly more often when AI was available. Judgment suspension rose slightly with incentives but stayed well below the no-AI control group. The incentives worked but were modest, as the researchers acknowledge. Whether larger or reputation-based incentives would close the gap further is unknown. In Study 4, the AI answer was shown to participants automatically without them having to ask. The researchers say this mirrors a growing reality where search engines display AI-generated summaries and writing assistants offer unsolicited suggestions. The effect barely changed. Without incentives, judgment suspension dropped from 35 percent without AI to 1 percent with automatically displayed AI answers. With incentives, it fell from about 39 to 7 percent. The researchers frame their results in the context of "Epistemia" , the tendency to accept AI answers because they sound convincing rather than actually checking them. A language model always has to produce an answer. It never pauses, even when it doesn't know something. When users hand off their judgment to such a system, they may adopt its lack of restraint, the authors write. The effect also runs counter to what "advice use" research has shown. People normally underweight outside advice and shift only about a third of the way toward an advisor's position. Participants in this study did the opposite, the authors say.