Auto-shown AI answers cut 'I don't know' from 35 percent to 1 percent in a movie-trivia study

Every condition offered an "I don't know" option. From Study 2 on it was priced: 10 cents for a correct answer, minus 10 cents for a wrong one, nothing for abstaining ^[9]. Under that rule an answer pays only when the chance of being right is above 50 percent ^[12]. Paying people helped a little. In Study 3 it cut AI requests from 5.27 to 4.53 out of six, and abstention rose slightly but stayed well below the no-AI group ^[11]. The accuracy figures in the report come from the unpaid arms, where people with AI access got 9.2 percent of questions right against 27.5 percent without it ^[7]. The AI group was right about a third as often ^[8].

Study 4 matches any tool that pre-fills a suggestion for a reviewer. The AI answer appeared without being requested, and the effect barely changed from the studies where participants had to ask for it ^[13]. The researchers say this setup mirrors AI summaries in search results and unsolicited suggestions from writing assistants ^[14].

The authors place the cause in the model's behaviour. They write that a language model always has to produce an answer and never pauses, and that people who hand it their judgment may adopt its lack of restraint ^[15]. The reported conditions do not include one where the model declined to answer or flagged low confidence. So the case for showing uncertainty inside the suggestion follows from the authors' explanation, and this study does not measure it. For a queue where a wrong approval costs more than a skipped item, I would build two things. One is a model allowed to return no suggestion. The other is a form that records the reviewer's own answer, abstention included, before the suggestion renders. The study tested neither.

The result also cuts against the advice-taking literature. People normally underweight outside advice and move only about a third of the way toward an advisor's position, and participants here did the opposite, the authors say ^[16]. They conclude that AI access turned some answers that would have been correct into errors ^[17]. Because the advice was mostly wrong, the pattern cannot be put down to sensible delegation to a reliable tool ^[20].

The figures describe a chosen workload. The authors picked film details because they rarely appear in online text and so invite hallucination ^[3], and Step 3.5 Flash was almost always wrong on them ^[1]. GPT-5.5, Claude 4.6 Sonnet and Gemini 3.5 Flash got most of the other questions right ^[4]. For the numbers to carry over to a review queue, the reviewer would need to know the item about as poorly as a viewer recalls a costume colour, and have about as little at stake. A reviewer whose name goes on the approval is risking more than a dime. The researchers acknowledge the incentives were modest and say the effect of larger or reputation-based stakes is unknown ^[19]. The Decoder notes that whether the deference holds as strongly outside movie trivia is still open ^[18].