Last year, a patient came through an emergency department for what the triage note described as lower abdominal pain, rule out appendicitis. She got a CT scan. The radiologist read the abdomen. The surgeons were called. The appendix was removed.
Nobody looked at the pancreas. Nobody was supposed to. The pancreas hadn’t been in question. But a year later we would find out that there was a nodule in it when the images were run through a new imaging and screening algorithm. That patient will be followed more closely now for screening and to prevent pancreatic cancer.
This is not a story about a missed diagnosis. It is a story about how medicine works. Emergency physician attention is a finite, directed resource, aimed at the presenting problem, the differential diagnosis, the plan for this patient at this time on this shift. The system functions precisely because clinicians focus on the task in front of them. Though we might expect them to manage every possible condition we have ever heard of, clinicians simply cannot focus on everything at all times.
But that doesn’t mean we shouldn’t do anything with data that has been collected, or the images that were taken. One of the key questions now is whether artificial intelligence may have a role to play in augmenting human focus.
I think it does. But not in the way most people are building it.
The dominant vision of AI in healthcare is a competitive one: a system that looks at the same data as the physician but tries to do the physician’s job better. Read the chest X-ray. Generate the differential. Suggest the diagnosis. Better, faster, cheaper. The implicit promise, and implicit threat, is substitution.
This vision has produced real tools, such as radiograph screening or ECG readers like the Queen of Hearts. It has also led to a decade of adoption failures, alert fatigue, and persistent skepticism from clinicians who have watched their workflows absorb new cognitive burdens in the name of efficiency. Having seen how detrimental prior technological promises have been, it’s no wonder the 11% adoption rate we have observed for ambient AI documentation in our own emergency department is not unusual. Most physicians credit the EHR—the last technology that was supposed to make our lives better—as the root cause of burnout in their practice. Physicians are not irrationally resistant to technology. They are rationally resistant to tools that compete for the attention they need to do their jobs.
The more interesting question is what happens when AI stops competing with humans for their attention and instead does something different?
In November 2024, a hospital in Ningbo, China, activated a deep learning system called PANDA (Pancreatic Cancer Detection with Artificial Intelligence), developed by Alibaba’s DAMO Academy. The tool does not wait for a physician to order a pancreatic workup but instead runs in the background on routine non-contrast CT scans ordered for entirely unrelated reasons: kidney stones, chest symptoms, trauma surveys, and the like. PANDA looks at the pancreas whether or not anyone has asked it to do so.
By early 2026, PANDA had reviewed more than 180,000 scans and flagged more than 20 pancreatic malignancies, 14 of them early-stage. Pancreatic ductal adenocarcinoma has a five-year survival rate of roughly 13%, almost entirely because it is typically found late, when tumors are inoperable. Stage I disease, however, is curable. It isn’t that this type of cancer is only findable at later stages. Rather, it becomes found at later stages, because nobody went looking earlier.
Here, the honest accounting gets complicated—and it should. PANDA flagged approximately 1,400 patients in Ningbo, of whom physicians determined only about 300 warranted further workup. That is a false positive rate that would be disqualifying in a frontline diagnostic tool. But as an opportunistic screening layered on scans already acquired, generating downstream referrals rather than immediate clinical actions, the calculus is different. The upstream cost is near zero; the downstream burden is real but manageable. But “manageable” is a judgment call, not a given, and it depends on a health system’s capacity, specialist availability, and patient tolerance for uncertainty, all of which vary enormously across settings. The false positive problem does not kill this idea. It defines the design constraints.
In our own emergency department, we have been studying a structurally similar phenomenon. Atrial fibrillation affects roughly 37 million people worldwide and is a leading cause of ischemic stroke. It is also intermittent. A patient who comes to the ED due to a fall, or shortness of breath, or a UTI, and happens to be in AFib during their stay, may convert to sinus rhythm before discharge and leave without a documented diagnosis, without anticoagulation, and with a substantially elevated stroke risk that no one addressed, because no one caught it.
In a retrospective cohort of 65,244 consecutive adult patients receiving continuous ECG monitoring at our academic ED, we found that 1,945 patients (3% of those monitored) were discharged with undiagnosed atrial fibrillation despite detectable AF on their continuous monitor. Of those, 71% met criteria for anticoagulation. Patients with undiagnosed AFib had an adjusted hazard ratio for subsequent ischemic stroke of 1.32 (95% CI 1.02-1.71) compared to patients with known AFib, a gap that was entirely preventable. The data was there. The signal was in the waveform the entire time only no one had been assigned to look for it.
The equity dimension of these findings deserves its own reckoning. Compared to patients with known AFib, those discharged with undiagnosed AFib were more likely to be younger, female, on Medicaid, without a primary care provider, and to identify as Black or Hispanic/Latino. Opportunistic screening failures are not distributed randomly. They fall along familiar fault lines.
Opportunistic screening, meaning AI that continuously reviews monitoring data for findings outside the presenting complaint, is not a new clinical idea. It is the logical extension of what a thoughtful physician would do if they had infinite attention. They don’t. The AI does.
Penda Health, a network of primary care clinics in Nairobi running more than 1,000 patient visits per day, has deployed a GPT-4o-powered tool called AI Consult inside its electronic health record. The design is instructive. The tool does not interrupt the clinician mid-history. It runs asynchronously during key decision points (diagnosis entry, treatment selection) and activates only when it detects a potential error, surfacing a color-coded alert: green for no concerns, yellow for moderate flags, or red for safety-critical issues like an inappropriate antibiotic or a missed comorbidity.
In a study of 106 clinicians across 15 clinics over three months, AI Consult was associated with a 32% reduction in history-taking errors, a 16% reduction in diagnostic errors, and a 13% reduction in treatment errors. The tool is projected to avert approximately 22,000 diagnostic errors and 29,000 treatment errors annually at Penda’s scale. Critically, it achieved this without requiring clinicians to change how they see patients. The human workflow has remained intact as the AI runs alongside it, watching for the things humans reliably miss owing to cognitive load, time pressure, and volume of patients and tasks.
There was also another unexpected finding. As the study progressed, clinicians in the AI group triggered fewer red alerts, suggesting the tool did not just catch errors but may have been teaching, as well.
That finding cuts both ways. If AI can accelerate clinical learning, it should be part of the conversation about medical education and quality improvement. But the inverse question shadows every example in this piece: what happens to a physician’s capacity to notice things (intermittent arrhythmias, incidental lesions, treatment mismatches) when the machine takes over the noticing? We have reasonable evidence for short-term error reduction, but there is almost no evidence regarding long-term skill trajectories. The capability atrophy we have documented in other AI-assisted clinical tasks may or may not apply here. We don’t yet know.
The best AI in healthcare is not the AI that competes with physician judgment. It is the AI that occupies the space physician judgment cannot occupy, providing a second read on an organ nobody ordered, the arrhythmia that appeared and disappeared between assessments, the antibiotic that was reasonable in isolation but wrong in context.
These three examples share a design philosophy that the field has not identified cleanly enough: a pancreatic cancer detector running on incidental CTs, a continuous ECG monitor watching for intermittent arrhythmias, a clinical decision support system that waits for you to finish before it speaks.
But this framework comes with its own set of obligations. Deskilling is an issue and must be studied. Likewise, the false positive burden is real and must be designed against. There is real harm when people are put under undue stress from potential “incidental” findings. AFib may convert before discharge. A nodule may turn out to be nothing. An antibiotic could be prescribed when there is no infection.
Medicine has long had a word for the moment when something is found because you were already there but looking at something else. We call it incidental. The word implies happenstance. AI can make that happenstance structural. They say, “Chance favors the prepared mind.” Whether we are prepared for these tools to make chance commonplace is open question.