Advice on a Barrier: Releasing AI Suggestions Only After Independent Human Work

When clinicians see an AI suggestion before they have formed their own view, they anchor on it. Expert pathologists overturned a correct judgment for a wrong AI estimate in about 7% of assisted decisions. Endoscopists shown computer-drawn boxes around suspected polyps narrowed their visual search and spent roughly half their inspection time looking at false alarms. The usual fix is to delay the suggestion. Fixed delays are crude, though: they waste time for a fast reader and are too short for a slow one, and in practice they create friction that users push back against. We propose instead to hold the AI output in a buffer and release it only when the human has visibly done independent work. That can mean committing an initial estimate, or, for live video where nothing gets "committed," having their gaze sweep the regions an expert would normally check. The central test compares this against a fixed timer tuned to the same average delay.

The design borrows a lesson from many-core computer architecture. There, processor cores that wait on a fixed schedule or keep re-checking a shared flag ("polling") waste time and slow unrelated work. A barrier (a point where every participant must arrive before any proceeds) triggered by an event removes most of that cost. Architects also measure the minimum amount of independent work that makes a synchronization worth paying for. We carry both ideas over to the human–AI handoff: a cognitive barrier on human events, and a measured cognitive synchronization-free region, the least independent work a reader must do before advice arrives for the bias rate to fall to near the unaided level. We are explicit that the hardware direction supplies the design principle and a latency budget to check, not custom silicon. One experiment below asks whether anything faster than display-frame timing matters at all.

Core Claim

Automation bias depends less on how long advice is withheld than on whether its release is conditioned on the reader's own work. Event-gated release should lower the rate of negative consultations and the severity of errors in those cases more than a fixed timer with the same mean delay, at equal or lower time per case. The comparison with a matched fixed timer is what makes this claim new. Delaying AI output is a known cognitive-forcing idea, and it is known to cause friction. Nobody in the literature we found has separated the effect of what triggers release from the effect of how long release is delayed. Nobody has built gaze-contingent release into a live detection pipeline either. The research file lists this as an open gap: eye tracking so far has only been analyzed after the fact, never used to control the AI.

Grounding in Prior Work

The motivating evidence comes from two groups. Rosbach and colleagues ran a factorial study of 28 pathology experts estimating tumor cell percentage: Automation Bias in AI-Assisted Medical Decision-Making under Time Pressure in Computational Pathology and Stuck on Suggestions. They found 38 negative consultations in 560 assisted decisions. A 10-second countdown did not change how often bias occurred, but it raised error magnitude in bias cases from 19.42 to 27.79. Reliance on advice, measured by the Judge Advisor System (JAS) score, rose from 0.49 to 0.55, and the AI advice carried an anchoring coefficient of 0.44. These studies vary pressure but always show advice immediately. We keep their task and change only when advice is released.

For endoscopy, Troya et al. showed that a detector finding polyps in 1.16 s gave no team speedup: the human with the detector took 2.90 s against 2.97 s unaided. Eye travel shrank, and false-positive misreadings rose from 4 to 6. Zhu et al. released eye-tracking data showing that experts scan in a "bottom U" pattern and check the periphery ring. Both papers treat the detector as a black box with fixed timing. We turn the scanning pattern they documented into the release condition.

The strongest algorithmic competitor is Brown et al.. They use an auditor model to hide predictions it judges unreliable, which improved team accuracy over humans alone in simulations on health-record data. That same paper constrains our design: when the human outperforms the model, hiding bad predictions stops helping. Hiding decides whether advice appears; our barrier decides when, so the two can be combined.

The hardware papers supply the method analogy. The Synchronization and Communication Unit of Glaser et al. replaces a 176-cycle software barrier with a 6-cycle event-driven one and cuts the minimum worthwhile synchronization-free region from about 1,700 to 42 cycles. LRSCwait/Colibri shows that polling slowed unrelated work to 26% of its normal speed, while a hardware queue caused no slowdown. It also points out that such queues need timeouts in case a participant hangs. The DPU's single-cycle global barrier (Shah et al.) is the same idea at larger scale. We reuse three things from this work: event-triggered release instead of polling or fixed waits, the synchronization-free-region metric, and the timeout, which becomes our safety bound. We do not claim that cycle-level latency matters to a human.

Research Program

We build one release controller. It sits between the model's output and the display and holds each suggestion until one of three conditions is met.

The first is a commit barrier. For static images such as pathology patches, the reader enters an initial estimate before the AI estimate and its cell detections appear.

The second is a gaze barrier, used for both static images and video. The eye tracker streams fixations to the controller. For a pathology patch, a suggestion is released once the reader's fixations have covered a set fraction of the tissue. For colonoscopy, a detection box is released once the endoscopist has fixated on the peripheral ring and the lower region within a short trailing window, following the expert pattern in Zhu et al.

The third condition is a timeout that releases the suggestion regardless, so that a true polyp is never hidden indefinitely.

The coverage threshold is the dose variable. Let $b(w)$ be the negative-consultation rate when advice is released after $w$ seconds of independent work, and let $b_0$ be the rate at which a reader's own correct call is changed by an unrelated perturbation, which we estimate from repeat unaided trials. The cognitive synchronization-free region is then the smallest $w$ such that

$$ b(w) - b_0 \le \epsilon , $$

where $\epsilon$ is a tolerated excess bias rate, set before the study to 1 percentage point. This mirrors how the hardware papers fix a 10% overhead tolerance. We estimate $b(w)$ with the linear mixed-effects models Rosbach et al. used, with random effects for reader and case.

Proposed Experiments

Pathology study. We reuse the tumor-cell-percentage task: FCOS cell detection on patches from BreCaHad, the colorectal set, and BreastPathQ. We enlarge the case pool from BreastPathQ's 2,395 patches. The original study had only 38 bias events, so we seed wrong AI estimates in a fixed, disclosed-afterward fraction of trials to obtain enough events for statistical power. Readers are pathologists and residents in a within-subject design with the two-week washout used by Rosbach et al., run both untimed and under the 10-second countdown.

Colonoscopy study. We use Zhu et al.'s 60 retrospective clips (41 with polyps, 19 without) and a 30 Hz eye tracker. Endoscopists are grouped by experience, because Stuck on Suggestions reports that experience reduces undue reliance.

Outcomes. For pathology we measure the negative-consultation rate, error magnitude within bias events, JAS score, anchoring coefficient, and time per case. For colonoscopy we measure reaction time, the share of time spent fixating false-positive boxes, false-positive misreadings, eye travel distance, and missed polyps. We also record whether readers try to bypass the gate, as a measure of friction.

Baselines

Immediate display. This is the naive reference point and matches current practice. On pathology it reproduces the published figures: a negative-consultation rate of about 7% (38/560) and bias-case error of 19.42 untimed and 27.79 under pressure, from Rosbach et al.. On colonoscopy the reference is 2.90 s assisted reaction time and about 50% of inspection time on false-positive boxes, from Troya et al.. Beating it means a lower bias rate without a worse miss rate.

Matched fixed timer. This is the ablation that isolates the key ingredient. Each reader's timer is set to the mean release delay that reader experienced under the event gate, so the average delay is identical and only the trigger differs. The event gate beats it if it gives a lower bias rate at the same or lower time per case. If the timer does as well, the core claim fails.

Auditor suppression. This is the strongest existing method. We implement Brown et al.'s approach with a logistic-regression auditor trained on the same model outputs and tuned on held-out cases with the same budget we use to tune our gate threshold. The file reports only significance for this method, $P < 8.2 \times 10^{-6}$ for team AUROC over human alone on VUMC and MIMIC-IV, and no reported result found on Lacuna for auditor suppression on pathology or endoscopy tasks, so we measure it ourselves. We also run auditor plus gate.

Why not just…?

Why not just add a software delay in the user interface? This is the right implementation and we use it. The controller is ordinary software acting on eye-tracker events at frame rate. The real question is whether a plain timer is enough, and the matched-timer arm answers it directly. If the timer performs as well as the gate, we will report that event triggering adds nothing.

Why not build dedicated synchronization hardware, as the architecture papers do? The time scales differ by roughly six orders of magnitude: 6-cycle barriers against 33 ms eye-tracker samples. We test this directly by injecting 0, 50, and 200 ms of random jitter into release times. If behavior is unaffected below about one frame, which we expect, custom silicon is unjustified for this problem, and we will say so. The hardware work then contributes the design principle and the timeout, not the chip.

Why not just suppress bad predictions with an auditor? Suppression cannot help when the model is right and the reader anchors anyway. By Brown et al.'s own account, it also stops helping when the human outperforms the model. It is a real competitor on cases where the AI is wrong, so we report results separately for cases where the AI is correct and cases where it is wrong. The gate should win mainly on the former, and on the latter only among expert readers.

Why not just train clinicians better? Board-certified experts still showed a 7% bias rate, and experienced endoscopists still fixated on false boxes. Training is cheap, however, so we include a briefing-only arm, a short automation-bias briefing before immediate display, to measure how much it closes the gap.

Decisive Checks

The proposal succeeds if the gate beats the matched timer on negative-consultation rate at equal or lower time per case, if bias-case error under time pressure falls below the immediate-display level, and if the timeout keeps polyp misses from exceeding immediate display. It also must produce a stable, estimable cognitive synchronization-free region per task. It fails if the matched timer ties the gate, and that outcome would be informative in its own right.