Does time pressure make doctors follow wrong AI advice?
In an unreviewed experiment with 155 German physicians, a 27-second limit increased movement towards incorrect sepsis advice and reduced doctors’ ability to distinguish it from correct advice. The study used simulated cases, not live care.
The short version
- What happened
- The within-participant experiment assigned 155 physicians ten sepsis vignettes: five under a 27-second limit and five with 146 seconds. Eight AI recommendations were correct and two were incorrect.
- Who it affects
- Clinicians and patients need AI interfaces that leave enough time and evidence for a recommendation to be challenged. This experiment does not prove that a deployed system will cause harm, but it shows why oversight should be tested under the conditions in which people actually work.
- What remains unproven
- Unreviewed preprint using ten simulated cases rather than clinical care.
Keep in mind
Unreviewed preprint using ten simulated cases rather than clinical care.
Only two vignettes contained incorrect recommendations, limiting separation of case difficulty from advice correctness.
The 27-second limit was deliberately severe, and timed trial completion was substantially lower.
Overall final accuracy did not differ; the correctness-specific result should not be turned into a blanket claim that time pressure made every decision worse.
Intensive-care experience was concentrated at the lower end and did not establish a protective threshold.
Source analysis · 1 linked source. Primary analysed source date: 29 October 2026.
Read the full report · 5 min The evidence, methods, implications and limitations
The measured problem was selectivity, not blanket reliance
A web experiment with 155 physicians found that severe time pressure made them more likely to move their sepsis estimates towards incorrect AI advice. Each doctor reviewed ten German-language clinical vignettes built from MIMIC-III intensive-care data, first estimated the chance of sepsis, saw a model estimate, then revised or retained the judgment. Five cases had a 27-second limit for the whole sequence and five allowed 146 seconds. The order and case allocation were randomized within participants. Eight recommendations were correct; one false positive and one false negative made up the two errors.
The study measured how much of the distance between a doctor’s first estimate and the model’s estimate was adopted. Time pressure increased uptake overall by 4.8 percentage points. That change came mainly from wrong advice: uptake rose by 0.146 on the zero-to-one scale (P < 0.001), while the 0.021 change for correct advice was not statistically clear (P = 0.272). On wrong-advice trials, moves at least halfway towards the model rose from 29.6% with more time to 50.8% under pressure. Among the 58 doctors with observations in every condition, the gap between uptake of right and wrong advice narrowed by 0.117 (P = 0.024).[1]
Wrong advice mattered, but the headline accuracy result needs care
Across all cases, final accuracy did not clearly differ between the two time conditions (P = 0.393). The more useful distinction appeared when the researchers separated correct and incorrect recommendations. Final estimates were less accurate under pressure on the two wrong-advice cases, with a 0.080 increase in Brier loss (P < 0.001); lower Brier loss is better. A baseline-adjusted analysis still found a difference, although doctors’ initial estimates on those cases had already been less accurate under pressure. The change from each doctor’s initial to final estimate was not clearly different between time conditions, so the experiment does not show that AI alone caused the full accuracy gap.
Experience offered no simple safeguard. The 153 doctors reporting intensive-care experience ranged from none to 37 years, but 45% had less than one year. Less experience was associated with greater advice uptake overall (Spearman’s rho -0.307, P < 0.001). It did not measurably change the time-pressure effect. For hospitals, the practical signal is that a nominal human override is not the same as meaningful oversight: interface tests should include wrong recommendations, real workflow constraints and enough time to inspect disagreement, especially for less experienced users.[1]
The distinction that matters
What changed when time was short
The experiment separates acceptance of correct advice from acceptance of incorrect advice; the difference between them is the safety-relevant result.
Correct advice
27 seconds versus 146 seconds
Uptake change +0.021; not statistically clear (P = 0.272)
Incorrect advice
Two error cases across the experiment
Uptake change +0.146; P < 0.001
All final estimates
Correct and incorrect advice combined
No clear accuracy difference (P = 0.393)
This is early evidence from a deliberately harsh simulation
The 27-second limit was calibrated to create pressure, not to reproduce a normal bedside assessment. Participants completed 1,346 of 1,550 assigned trials; completion fell from 98.2% with 146 seconds to 75.5% under pressure. The vignettes had an artificial 50% sepsis prevalence, doctors were told the model was about 80% accurate, and only two cases contained wrong advice. Case difficulty therefore cannot be cleanly separated from advice correctness, and results conditional on completed trials may overstate uptake if rapidly accepting advice helped a doctor finish.
The paper is a medRxiv preprint and has not been peer reviewed. Recruitment through German professional societies and one academic hospital may have favoured physicians interested in AI. A stronger assessment would use more varied errors, balanced workflow sequences, realistic team interruptions and prospective observation in clinical environments, while measuring patient-safe actions rather than probability estimates alone. The authors reported no specific funding or competing interests; participant-level data are promised after peer-reviewed publication, while analysis code is available only on request.[1]