How the AI study measured willingness to suspend judgment
To suspend judgment means to hold back an answer when there is not enough to go on, which in this study meant choosing "I don't know". This study of overreliance on AI measured something that comes before accepting a wrong answer: whether people still use the option to decline once an AI assistant is in the room. The study found that people with AI on hand almost stopped using that option.
The AI advice study at a glance
The study ran five experiments with 3,132 people, and four of them were preregistered. The key facts about the paper are listed below.
Study overview
Preregistered means the researchers published their hypotheses and analysis plan before collecting any data. A direct replication repeats an earlier experiment to check that the result holds a second time.
Suspend judgment rates with and without AI advice
Every participant got six difficult questions about films and could decline any of them. The table shows the share of answers declined in each experiment, with and without AI advice available. Studies 2 to 4 also had a condition with money at stake, so the figures here are the ones without it.
| Experiment | Participants | No AI | AI available |
|---|---|---|---|
| Study 1a | 314 | 0.36 | 0.06 |
| Study 1b | 310 | 0.44 | 0.03 |
| Study 2 (no stakes) | 812 | 0.17 | 0.01 |
| Study 3 (no stakes) | 843 | 0.32 | 0.02 |
| Study 4 (no stakes, advice shown automatically) | 853 | 0.35 | 0.01 |
The rate fell in all five experiments. The largest gap was in Study 1b, where declining dropped from 0.44 to 0.03. In that experiment the AI's wrong answers were written in advance and then displayed.
"In five experiments (N = 3,132; four preregistered, one direct replication), participants answered difficult questions and could always decline to respond." (Abstract) ／ "Participants in the baseline condition were substantially more likely to suspend judgment than participants who could seek AI advice (0.36 vs. 0.06; one-tailed t-test: t = 8.295, p < .001; see Figure 1a)" (Study 1a) ／ "Participants in the baseline condition were again substantially more likely to suspend judgment than participants who could seek AI advice (0.44 vs. 0.03; t = 11.772, p < .001; see Figure 1b)" (Study 1b) — from the arXiv paperView official source →
Why trust in AI advice cannot explain the drop
One possible explanation for this kind of result is reasonable trust: people let the AI answer because it knows more. The design of this study rules that out from the start, because the advice was always wrong.
The AI advice was built to be wrong, isolating AI accuracy
The researchers picked questions the AI would get wrong, on purpose. This separates two things that usually come together: using AI, and the AI being right. Whatever made people stop declining, it was not good advice.
The authors suggest that AI suggestions may shift the point at which people decide they know enough to answer. In everyday terms, a ready-made answer may make it harder to say "I'm not sure".
"We engineered the questions so that AI advice was wrong, separating AI use from its accuracy." ／ "As AI suggestions grow ubiquitous and unsolicited, they may not simply affect answer accuracy; they may even alter the metacognitive threshold at which people decide whether they know enough to answer." (both from the Abstract) — from the arXiv paperView official source →
AI accuracy fell to a third while AI confidence rose 2.5x
Across the studies without money at stake, people without AI got 27.5% of answers right, and people with AI got 9.2% right. That is about a third as often.
Their confidence moved the other way: the mean score went from 29.6 to 75.9, about 2.5 times higher. So people with AI answered more questions, got fewer of them right, and felt much surer of themselves.
"in the absence of monetary incentives, participants answered more questions but were correct about a third as often as when AI was unavailable (pooled correctness: 27.5% vs. 9.2%), while confidence was roughly two and a half times as high (mean confidence: 29.6 vs. 75.9). Without incentives, AI access made people far more assured and far less accurate." (Discussion) — from the arXiv paperView official source →
What AI incentives changed about AI dependence
Studies 2 to 4 added money to the task. Each correct answer earned $0.10, each wrong answer lost $0.10, and declining cost nothing.
The researchers had predicted in advance that money would bring back declining, and that AI would weaken that effect. Money helped only a little, and the AI groups still declined far less than the groups without AI.
Stakes raised suspend judgment rates, but not to baseline
Money made people decline a little more often, but nowhere near the level seen without AI. In Study 3, the AI group went from 0.02 to 0.08. The group without AI went from 0.32 to 0.41.
The prediction about AI weakening the effect of money was not supported in any of Studies 2, 3 or 4. AI and money worked mostly on their own. AI cut declining sharply whether or not money was involved, and money raised it slightly whether or not AI was there.
"Across Studies 2–4, then, our pre-registered prediction of a negative AI × stakes interaction on suspension was not supported (p = .274, .506, and .784, respectively). AI availability and stakes acted largely independently: AI sharply reduced suspension regardless of stakes, while stakes modestly increased suspension—most clearly in Study 3—regardless of AI." (Study 4) — from the arXiv paperView official source →
Incentives raised accuracy only with AI advice available
Money improved accuracy only for people who had AI. In Study 3, the AI group rose from 0.11 to 0.16, while the group without AI stayed at 0.28 and 0.27. Study 4 showed the same pattern: 0.07 to 0.14 with AI, and a flat 0.27 without.
With money at stake, people also asked the AI less often. Requests for advice fell from 5.44 to 4.93 out of six questions in Study 2, and from 5.27 to 4.53 in Study 3.
Money did not make people better judges. It made them lean on the AI less.
"Like Study 2, when AI advice was available, monetary stakes increased correctness (0.11 vs. 0.16; b = 0.055, SE = 0.017, t = 3.164, p = .002). By contrast, when AI advice was unavailable, stakes had no effect on correctness (0.28 vs. 0.27; b = -0.009, SE = 0.022, t = -0.382, p = .703; see Figure 3b)" (Study 3) ／ "Participants sought AI advice substantially less frequently when stakes were present than when they were absent (4.53 vs. 5.27 times out of six questions; b = -0.744, SE = 0.182, t = -4.083, p < .001; see Figure 3c)" (Study 3) — from the arXiv paperView official source →
Automatic AI advice still cut the will to suspend judgment
The paper notes that AI suggestions are becoming ubiquitous and often show up unasked. A search page can open with an AI summary, and a writing tool can suggest the next sentence.
Study 4 tested that setup. Advice that appeared without a request cut declining just as much.
Study 4 put the AI advice on screen without a request
Study 4 (N = 853) was the same as Study 3, except that the AI advice appeared on its own instead of on request. Declining still fell from 0.35 to 0.01 without money at stake, and from 0.39 to 0.07 with it.
People did not need to ask the AI for the effect to appear. Seeing the answer was enough. That setup is close to tools where an AI summary is switched on by default.
"Study 4 (N = 853) therefore removed participants' control over whether to receive AI advice. It was identical to Study 3 except that, in the conditions where AI advice was available, it appeared automatically rather than on request." ／ "Access to AI advice continued to dramatically reduce judgment suspension, both when monetary stakes were absent (0.35 vs 0.01; b = -0.333, SE = 0.029, t = -11.56, p < 0.001) and when they were present (0.39 vs 0.07; b = -0.322, SE = 0.032, t = -10.18, p < 0.001)." (both from Study 4) — from the arXiv paperView official source →
Epistemia and metacognition: willingness, not capacity
Metacognition is the ability to judge what one knows and does not know. The authors argue that AI does not damage this ability, because money brought part of it back.
If people had lost the ability, a small payment would not have changed anything. What changes is their willingness to use it.
The authors also connect the pattern to epistemia. The term describes accepting AI output because it sounds plausible and reads smoothly, without checking it. Questions about who should take credit for AI-assisted writing are covered in "AI Written Articles Under a Real Writer's Byline".
The per-experiment results sit in the paper's PDF, spread across sections and figure notes. When you want to keep each figure next to the study and condition it came from, try the tool that converts a PDF to Markdown and keeps its headings, tables and lists.
The study points to overreliance on AI as a change on the human side, not only a question of how accurate the model is. With a fluent answer in view, people stop saying "I don't know", get fewer answers right, and feel more sure.
Money for correct answers brought some of the caution back, which the authors read as a sign that the habit can be partly corrected. In practice, that argues for keeping a step in the work where "I don't know" is still an accepted answer, and where wrong answers carry a real cost.
"the near-elimination of suspension under AI access is not a fixed cognitive limitation but a contextual response that monetary consequences partly correct: people can recruit their own judgment when motivated, so what AI changes is willingness, not capacity." ／ "These findings can be read through the notion of epistemia … : the tendency to accept AI outputs for their surface plausibility and syntactic coherence rather than through verification." (both from the Discussion) — from the arXiv paperView official source →
FAQ
- How many people took part in the AI advice study?
- 3,132 people across five experiments. Four of the experiments were preregistered and one was a direct replication. In every one, participants answered six hard questions about films and could decline any question.
- Was the AI advice in the study correct?
- No. The researchers chose questions the AI got wrong on purpose. That way the result shows what having AI does to people, apart from whether the AI is right, so the drop cannot be explained as handing work to a good tool.
- Does AI advice make people less accurate?
- In this study, yes. With no money at stake, people with AI available were right on 9.2% of answers, against 27.5% without AI, so about a third as often. They also answered more questions, because they stopped declining.
- Do incentives reduce overreliance on AI?
- Only partly. When correct answers paid and wrong ones cost money, people asked the AI less, answered more accurately, and declined more often. They still declined far less than people who had no AI at all.
Related articles
AI Streamer Faces: Too Perfect Loses EngagementA Journal of Business Research study of 347 AI streamers ties asymmetric faces to more engagement and typical faces or intense expressions to less.
Claude Fable 5.2: 20% price cut, up to 30% fasterAnthropic released Claude Fable 5.2 on November 9, 2026. API prices fall 20% to $8/$40 per million tokens, and output is up to 30% faster than Fable 5.1.
PayPay security and passkeys in 10 Japan pay appsPayPay security compared with 9 other Japanese payment apps: passkeys, extra checks to send or pay, and fraud compensation deadlines.
AI fact check: only 40% of cited stats hold upAI fact check: Trakkr traced 1,285 stats cited by ChatGPT, Perplexity, Gemini, Claude and AI Overviews. Only 40% led to a matching study. How to check one.