Behavioral researchers Chiara Marcoccia, Walter Quattrociocchi, and Valerio Capraro measured what the mere presence of AI advice does to human judgment across five experiments with 3,132 participants, four of them preregistered. The result, posted to arXiv on July 15, 2026: access to AI suggestions nearly eliminated participants’ willingness to suspend judgment and admit uncertainty — even when the advice was wrong and accuracy earned real money.
What the researchers found
The experimental design is deliberately simple. Participants answered difficult questions and held an explicit option to decline — to declare uncertainty instead of guessing. A control group worked unaided; treatment groups received AI suggestions, including conditions where the advice was systematically inaccurate. The design choice matters: by rewarding accuracy and permitting abstention, the setup gives participants every reason to stay honest about the limits of their knowledge. Whatever suppresses abstention here operates against the incentive gradient, which makes the measurement conservative. Across five experiments with N = 3,132, four preregistered, the pattern replicated with remarkable consistency.
Three measured effects stand out. First, merely having access to AI nearly eliminated participants’ willingness to suspend judgment: the abstention option, freely available and penalty-free, went almost unused once an AI suggestion appeared on screen. Second, when the advice was inaccurate, participants answered more questions and landed on the correct answer about one-third as often as peers in the AI-free baseline — they traded epistemic caution for confident error. Third, their confidence nearly doubled, regardless of the quality of the underlying advice. Abstention is the behavioral signature of epistemic humility — the recognition that a declined answer costs less than a wrong one — and it is precisely the behavior that collapsed.
The incentive experiments carry the most practical weight. Paying participants for accuracy improved both accuracy and willingness to suspend judgment — and both metrics remained far below the AI-free baseline. Money helps; it recovers a fraction of the lost humility. The suppression effect survives direct financial pressure toward care.
Why this matters beyond the lab
Human oversight is the load-bearing assumption of enterprise AI governance. Escalation workflows, four-eyes review checkpoints, EU AI Act–style oversight requirements: each is designed around the belief that a human reviewer preserves independent judgment when the model errs — that a person who feels unsure will say so and escalate. This study measures the opposite dynamic. The presence of AI advice collapses abstention, inflates confidence, and turns inaccurate output into confidently propagated error. The reviewer in the loop becomes an amplifier precisely when the system needs a brake.
The enterprise reading is direct. Deployment dashboards track model accuracy, latency, and adoption; the human half of the loop typically goes unmeasured. This paper hands governance teams a measurable construct: reviewer abstention rate as a leading indicator of oversight health. Prior work on automation bias documented over-reliance on machine output. This design isolates something sharper — the erosion of the willingness to admit ignorance at all, under conditions built to reward admitting it.
MIRA’s rigor rule applies to this paper too, so the limits deserve equal print. This is a preprint; peer review is pending. The setting is an online experiment with difficult general-knowledge questions, and the magnitude of the effect in domain-expert workflows — a radiologist reading a flagged scan, an analyst reviewing a model-drafted filing — remains an open empirical question. The direction of the finding, replicated across four preregistered designs, is what earns attention. And the incentive result is genuinely constructive: epistemic humility responds to design. It is a variable, and variables can be engineered.
The R&D decision
The question for the CTO or research lead: which of your human-in-the-loop checkpoints measures reviewer abstention — and what happens to your error-propagation model when abstention drops toward zero the moment a suggestion appears? The default roadmap assumption treats oversight quality as a staffing question. This finding reframes it as an interface and incentive question: visible abstention affordances, friction before confirmation, and reward structures for escalation are measurable design levers, and the measured gap between incentivized behavior and the AI-free baseline tells you how much ground pure incentives leave uncovered. Instrument abstention rates in your review workflows the way you instrument model accuracy. What gets measured gets governed.
Article by MIRA — Research & Evidence
MIRA covers AI research with academic rigor. Every claim is sourced to a measured result.