AI access cuts 'I don't know' responses from 44% to 3%, study finds

A study of more than 3,000 participants found that access to an AI assistant cut the share of respondents willing to say "I don't know" from 44 percent to 3 percent. The AI was wrong on nearly every question, yet participants with access reported higher confidence and produced correct answers at roughly one-third the rate of their unassisted peers. The 41-percentage-point swing in abstention (the deliberate choice not to answer) is the headline figure in a result that complicates the productivity argument for AI assistants.

The experiment, reported by The Decoder, isolated AI access as the only variable between two matched groups. Both groups faced the same questions. Only one could consult the model. Researchers measured three dependent variables: willingness to admit ignorance, self-reported confidence on a post-task survey, and factual accuracy against a known ground truth.

What did the 3,000-participant study measure?

The design targeted three measurable outputs. First, the rate at which participants said "I don't know" rather than guessed. Second, self-reported confidence on a post-task survey. Third, factual accuracy against a known ground truth.

The questions were chosen from a domain in which the underlying AI model performed poorly. The model returned incorrect or fabricated answers at a near-uniform rate across the question set. That selection ensured any accuracy gap would reflect user behavior rather than model capability.

In the control group without AI access, 44 percent of participants chose to abstain rather than risk a wrong answer. In the group with AI access, that figure collapsed to 3 percent. The 41-point shift is the single largest behavioral change documented in the study.

Participants with AI access also registered higher confidence on the post-task survey. The combination of increased certainty and decreased accuracy is the central paradox the data expose. Users did not become better informed. They became more willing to commit.

How much less accurate were the AI-assisted participants?

AI-assisted respondents produced correct answers at approximately one-third the rate of the unassisted group. The study design controlled for question difficulty, so the gap reflects behavior rather than the model occasionally helping.

That ratio implies a 67 percent reduction in accuracy attributable to AI access. For every three correct answers given without AI, the assisted group produced roughly one. The pattern suggests participants deferred to the model's output rather than verifying it against prior knowledge or external sources.

The authors describe the effect as near-total substitution of model judgment for human judgment. Users do not appear to have used the AI as one signal among many. They appear to have used it as the signal.

Why does confidence rise while accuracy falls?

The result points to a documented asymmetry in large language model outputs. Models produce fluent, declarative prose even when the underlying answer is wrong or fabricated. The fluency signal conditions users to suppress doubt, regardless of the underlying probability that the answer is correct.

When users cannot see model internals and receive no calibration cues (visible indicators of how confident the model actually is), they treat the response as authoritative. The data describe exactly this collapse: a 44-to-3 percent swing in abstention paired with a two-thirds reduction in accuracy.

The mechanism matters because abstention is a load-bearing element of decision-making in medicine, law, journalism, and policy. A workforce trained to defer to AI outputs on every query will abandon abstention at scale. The cost of that abandonment shows up as wrong answers delivered with high confidence.

The behavioral pattern carries a familiar name in the calibration literature: the automation complacency effect (the tendency for human operators to accept machine outputs without verification, especially under time pressure). The 3,000-participant study is one of the first large-N demonstrations of the pattern with consumer-grade AI assistants rather than industrial control systems.

What were the study's limits?

The source does not specify the recruitment pool, the fieldwork date, or the institution behind the work. The 3,000-participant figure refers to the combined sample across both groups. The questions used were selected to be adversarial to the model, which limits generalization to domains where AI performs well.

The study also does not report whether participants were told the model's accuracy rate before the task. A pre-task disclosure could plausibly reduce the deferral effect by reminding users that the model is fallible. The findings therefore describe behavior under one specific condition: access without calibration cues.

These limits do not undermine the central result. The 44-to-3 percent swing in abstention is large enough that confounders (unmeasured variables that could distort the result) would need to be implausibly large to explain it away. The 67 percent accuracy reduction is similarly robust.

What does this mean for AI deployment?

The result lands at a moment when enterprises are scaling AI assistants across customer support, legal review, and clinical documentation. The implication is direct: deployment without calibration scaffolding will degrade aggregate decision quality rather than improve it.

Two design responses follow from the data. First, products must surface model confidence to users rather than bury uncertainty in fluent prose. A numeric confidence score, a verbal hedge, or a refusal option can preserve the user's ability to abstain.

Second, interfaces must keep a low-friction path to "I don't know" so users do not feel compelled to commit to a model-generated answer. The 3,000-participant data show what happens when that path is closed by default: users stop saying "I don't know" and start being wrong.

The next research wave will test whether explicit confidence prompts or abstention buttons close the 67 percent accuracy gap. Product teams can run the same test inside any deployed assistant by A/B-testing (randomly assigning users to two variants and comparing outcomes) a confidence-disclosure variant against the current default. Until those results land, the 3,000-participant benchmark stands as the cleanest empirical case that AI access, absent calibration, makes users worse judges of what they do not know.