Study Finds AI Access Slashes “I Don’t Know” Responses While Boosting Overconfidence

1. What happened, who was involved, and when

A research team led by Matthias Bastian published new findings on September 26 2026 after running a series of five experiments involving 3,132 volunteers. The studies examined how the mere availability of a conversational AI influences people’s willingness to acknowledge uncertainty.

2. Core findings and experimental details

• Design: Participants faced trivia questions that required fine‑grained visual knowledge from movies (e.g., the color of a team’s uniform in Bend It Like Beckham). The primary model offered, called Step 3.5 Flash, was deliberately chosen because it was wrong on almost every such question, making the task a stress test for reliance on AI. Other models—GPT‑5.5, Claude 4.6 Sonnet, and Gemini 3.5 Flash—performed well on the easier items but also stumbled on the hardest ones.

• Effect on “I don’t know” responses: In the first two experiments (1a and 1b) participants could decide whether to consult the AI. Without AI access, they chose to withhold judgment on 36 % and 44 % of the items, respectively. When the AI was available, those rates collapsed to 6 % and 3 %.

• Confidence vs. accuracy: In Study 2, confidence scores (out of 100) jumped to 75.9 with AI access, compared with 29.6 when the model was unavailable—more than double. At the same time, correct‑answer rates fell from 27.6 % to just 10 %. Across all five studies, participants without incentives who could query the AI answered correctly on only 9.2 % of the questions, versus 27.5 % when the AI was blocked.

• Financial incentives: Studies 3–4 introduced a modest payoff (gain of $0.10 for each correct answer, loss of $0.10 for each wrong answer, and zero for “I don’t know”). The pre‑registered hypothesis expected incentives to encourage more “I don’t know” replies, but the interaction with AI availability was not statistically significant. Incentives did modestly curb AI queries—participants asked for help 4.53 times on average out of six possible requests in Study 3, versus 5.27 times without incentives. Accuracy improved slightly when incentives were present, but judgment suspension remained far below the no‑AI baseline.

• Automatic AI suggestions: Study 4 presented AI answers automatically, mimicking the way modern search engines and writing assistants surface AI‑generated content without prompting. The pattern persisted: judgment suspension fell from roughly 35 % (no AI) to 1 % (auto‑shown AI) without incentives, and from about 39 % to 7 % with incentives.

• Interpretation: The authors label the phenomenon “Epistemia,” the tendency to accept AI output because it sounds authoritative, even when the model has no mechanism to admit ignorance. This runs counter to classic “advice‑use” literature, where people typically weight external counsel conservatively and shift only a third of the way toward the advisor’s stance. Here, participants swung the opposite direction, sometimes turning a correct answer into an error by following the AI.

• Broader research: The paper cites complementary work, including a Swiss Business School survey of 666 respondents that found a strong negative link between AI reliance and critical‑thinking scores, and a Microsoft study reporting that brief AI‑assisted sessions can diminish problem‑solving persistence on subsequent tasks. Both suggest that AI use taxes metacognitive resources, though higher education levels may buffer the effect.

3. Industry context and implications

These results arrive as AI‑generated answers become embedded in everyday tools—search engines now surface large language model summaries, and writing assistants push unsolicited suggestions. The research warns that such ubiquity could erode the habit of admitting uncertainty, a cornerstone of sound decision‑making. While improving model accuracy remains a priority, the authors argue that preserving users’ willingness to say “I don’t know” may be equally critical for maintaining human judgment in an increasingly AI‑augmented world.