A preregistered study of 455 Prolific participants tested two AI advisor policies: advise on every decision, or stay quiet when the model's own confidence crosses a threshold. The same selective-abstention policy raised advice acceptance 15.1 percentage points for perceived experts (38.2% to 53.3%) and lowered it 5.3 points for everyone else (73.4% to 68.1%).
Why I kept re-reading this
The advisor policy is the user segmentation. Most product teams I work with treat it as a global design question. Ship on every decision, or ship only when the model is confident. Pick one.
The paper breaks that framing. Same product. Same accuracy. Two populations inside the same workforce. Opposite effect.
A preregistered mediation names the mechanism. Experts in the always-on condition pay a reconciliation cost on every piece of advice that disagrees with their judgment. The cost compounds across repeated tasks. By the end of a series of eight, their acceptance has quietly eroded. The quieter policy removes the cost, and their pattern stabilises.
The variable moving the outcome is accumulated evaluation burden. Accuracy stays constant across conditions.
What this looked like in a room
A VP of Operations I was coaching in September. Munich-based insurer, 2,000 adjusters, pilot rollout of an AI advisor on claims triage. Two policies on the whiteboard. The product manager wanted always-on because "more coverage equals more lift." The ops lead wanted selective because "senior staff keep muting the thing."
I asked how acceptance looked by tenure. Three-year adjusters: climbing. Twelve-year adjusters: flat, then falling off a cliff. The PM read it as resistance to change. The ops lead read it as the AI being wrong.
Both were wrong. The senior adjusters were paying a reconciliation cost on every piece of advice they already knew the answer to. Eight weeks in, they had stopped reading it. The juniors were still leaning on it, and taking the lift the AI was built to provide.
The fix was less advice for the people drowning in it. More advice for the people still learning the shape of the job. The debate had been framed as coverage versus noise. The actual axis was expertise.
The paper puts a number on what I had been calling a vibe. The senior population is subsidising the junior population's dependence on the advisor, and nobody in the product review has priced the subsidy.
"The same silence speaks differently to different users."
Ship both policies, route them by expertise, and stop treating the advisor as one product with one dial.