An LLM decision-support system ran inside a tertiary emergency department for four weeks. No adverse events. Reviewers rated 99 of 100 sampled outputs clinically appropriate. Clinician use fell from 68% to 30% — and the odds of opening it fell 28% for every additional hour of shift workload. Drag the hour. Watch the field go dark.
Primary source: “Large language model-based clinical decision support in the emergency department,” Nature Medicine, 19 August 2026 — a single tertiary emergency department
Workforce scale: CMS Medicare Physician & Other Practitioners — by Provider and Service, CY 2024 vintage, via MIMI Labs
Every argument about clinical AI is an argument about whether the output is correct. This study answered that question and then kept going, and the second answer is the one nobody budgeted for.
Accuracy wasn’t the failure mode. Attention was.
The paper reports one number that does all the work: an adjusted odds ratio of 0.72 per additional hour of shift workload (95% CI 0.62–0.83). Every hour deeper into the shift, the odds that a clinician opens the tool are multiplied by 0.72. That is the finding. What follows is that number drawn out to the length of a real shift.
One hundred decisions, at hour 0
Each square is one decision moment where the tool was available. Filled squares are the moments a clinician opens it, at the rate the published odds ratio implies for that hour of the shift. The curve underneath is the same thing, continuously, with the paper’s confidence interval shaded.
hour 0
68% (as reported)
Opens it
68%
at hour 0
Skips it
32
of 100 moments
Odds vs. hour 0
1.00×
0.72 per hour
Rated appropriate
99%
of 100 sampled outputs
100 decision moments — filled = tool opened
68
of every 100 moments the tool gets opened at the start of the shift.
The study’s own endpoints: adoption 68% → 30% across four weeks, zero adverse events, 99 of 100 sampled outputs clinically appropriate.
Drag the hour slider. The squares are not a simulation of individual people — they are the published odds ratio expressed as a rate out of 100.
Probability of use across a shift — OR 0.72/hour, 95% CI band
central estimate (OR 0.72)95% CI (0.62–0.83)beyond hour 8 — extrapolation
What the curve is, and what it isn’t
The paper reports an odds ratio. It does not publish this curve. The curve is what that odds ratio implies once you anchor hour 0 at the reported starting adoption rate — which is why the anchor is a slider and not a constant. Move it and the shape is unchanged; only the height moves. That is the honest content of an odds ratio: it tells you the slope, not the level.
80/20 — the part that transfers
If your tool’s value depends on someone choosing to open it, your evaluation needs a second endpoint next to accuracy: use, stratified by workload. Not blended. Not week one. A tool that is 99% right and opened at hour ten is a different product from the same tool opened at hour one, and only one of those is the product you piloted.
How far this has to travel
This was one emergency department, one country, four weeks. For scale: 48,681 individual emergency medicine clinicians billed Medicare Part B in 2024, inside a Part B workforce of 1,147,252 clinicians across 85 specialties. The bars below are Medicare encounter-days per clinician per year — the interquartile range, with the 90th percentile marked. Medicare is only a slice of any real panel, so every bar understates the actual day.
Medicare encounter-days per clinician, 2024 — p25–p75 bar, ◆ = p90
Emergency medicine sits low on this axis precisely because most ED patients are not Medicare beneficiaries. The spread within each specialty is the point: the median clinician and the 90th-percentile clinician are not having the same shift, and the odds ratio above says they will not make the same decision about the tool.
Where this is thin
One site. A single tertiary ED. The 0.72 is an association measured inside one staffing model, one EHR, one set of local norms. Nothing here establishes that it replicates.
Twelve hours is past the data. The confidence interval widens as you extrapolate, and the shaded region past hour 8 is drawn in grey for that reason. A 4% figure at hour 12 is arithmetic, not an observation.
“Clinically appropriate” is a low bar. 99 of 100 sampled outputs passed review. Appropriate is not the same as useful, and neither is the same as changed the plan. The study did not report an outcome the tool moved.
The squares are a rate, not people. Filled squares are drawn deterministically from the probability, so the field is a picture of a proportion — not a Monte Carlo of a hundred clinicians.
The CMS bars are the Medicare slice only. Encounter-days come from Part B claims for individual providers; commercial, Medicaid and uninsured volume is absent. Treat each bar as a floor.