Triage is the first decision in emergency care and arguably the most consequential. It decides who waits and who does not. It is usually made in a few minutes, by a nurse, with incomplete information, in a crowded room.
Artificial intelligence is being pitched as a way to make that decision faster and more consistent. Some people find that obvious. Others find it alarming. We think both reactions deserve a proper hearing, so this post makes the strongest case for each side, using evidence rather than hype or fear, and then leaves the verdict to you.
To be precise about the question: we are not asking whether AI should assist triage nurses, which is less controversial. We are asking whether an algorithm should be allowed to make the triage decision itself, assigning priority with humans monitoring rather than deciding.
The case for: human triage is already failing, and quietly
The strongest argument for AI triage is not that algorithms are brilliant. It is that the current system is worse than most people assume.
In 2023, JAMA Network Open published a study by Dana Sax and colleagues at Kaiser Permanente Northern California. They examined more than 5.3 million adult emergency department visits across 21 emergency departments between 2016 and 2020, all triaged using the Emergency Severity Index, the most widely used triage system in the United States. Using an algorithm to classify each visit, they found that mistriage occurred in 32.2% of encounters: 28.9% were overtriaged and 3.3% undertriaged.
The sensitivity of triage for identifying patients with high-acuity illness, those who went on to need a life-stabilising intervention, was 65.9%. Roughly one in three of the sickest patients was not placed in the top two categories.
The study also found troubling differences. In adjusted analyses, Black patients had an 18.5% greater relative risk of undertriage compared with White patients.
Proponents of AI triage draw a direct conclusion from evidence like this. Humans get tired, rushed and biased. They anchor on first impressions. They triage differently at 3am than at 3pm. An algorithm trained on outcomes could, in principle, apply the same standard to every patient, every time, and could be audited in a way individual judgement cannot. If the status quo misses a third of the sickest patients, the burden of proof should be on those defending it.
There is also a capacity argument. Emergency departments in many countries face staff shortages. If a safe algorithm can handle routine prioritisation, experienced nurses can spend more time on the patients whose presentations are genuinely ambiguous.
The case against: algorithms fail differently, and at scale
The strongest argument against is not that AI is inherently untrustworthy. It is that we have repeatedly deployed clinical algorithms that looked good on paper and performed poorly in practice, and that when an algorithm fails, it fails for everyone at once.
Consider sepsis prediction, a close cousin of triage. In 2021, JAMA Internal Medicine published an external validation by Andrew Wong and colleagues of the Epic Sepsis Model, a proprietary tool implemented in hundreds of US hospitals. Studying 38,455 hospitalisations at Michigan Medicine, they found an area under the curve of 0.63, well below the performance its developer had reported. Despite generating alerts on 18% of all hospitalised patients, the model did not identify 67% of the patients who had sepsis.
A model widely used in real hospitals missed two thirds of the cases it was built to catch, while alerting on nearly one in five patients. That combination, missed cases plus alert fatigue, is exactly what opponents fear in triage.
Bias is the second concern. In 2019, a study led by Ziad Obermeyer in Science examined a commercial algorithm used to identify patients for extra care management. The algorithm used health care costs as a proxy for health need. Because less money had historically been spent on Black patients with the same level of illness, the algorithm systematically underestimated how sick they were. At any given risk score, Black patients were considerably sicker than White patients. The authors, Obermeyer, Brian Powers, Christine Vogeli and Sendhil Mullainathan, calculated that removing the bias would raise the share of Black patients flagged for extra help from 17.7% to 46.5%. Algorithms built on this cost-based approach were, by their estimate, applied to around 200 million people in the US each year. The algorithm was not told anyone’s race. It learned the bias from the label it was asked to predict. Human bias can be retrained one clinician at a time. Algorithmic bias is automated, consistent and can be invisible until someone looks for it.
Opponents also raise accountability. If a nurse undertriages a patient, there is a professional responsible and a process for learning. If an algorithm does it, responsibility is diffused between the vendor, the hospital that bought it and the clinician who did not override it.
Finally, there is the problem of drift. Patient populations change, new diseases emerge, coding practices shift. A triage model validated in one hospital in one year can degrade silently elsewhere or later.
Rebuttal: the case for answers back
Each side has a reply to the other, and the debate only gets interesting at this stage.
Supporters of AI triage respond to the sepsis evidence by saying it proves the case for better validation, not against algorithms. The Epic model was proprietary, was deployed widely before independent testing and predicted a condition whose definition is itself contested. A triage model built transparently, tested prospectively and monitored locally would be a different object. They make a similar point about the Obermeyer study: the bias came from choosing the wrong target, cost instead of need. The researchers showed that changing the label could remove much of the disparity. A known, measurable bias can be fixed in a way that the scattered biases of thousands of individual clinicians cannot.
They also argue that the comparison should be with the real alternative, not an ideal one. The Kaiser Permanente study suggests human triage misses about a third of the sickest patients and undertriages Black patients more often. An algorithm does not need to be perfect. It needs to be better than that, and to be checked continuously.
Opponents reply that “we will validate it properly” is exactly what was said before previous deployments. The incentive structure has not changed: vendors want to sell quickly, hospitals want to cut costs and waiting times, and independent validation is slow and expensive. They also point out that the mistriage figures in the Kaiser study were themselves generated by an algorithm classifying outcomes after the fact. Retrospective labels of “correct” triage are not the same as knowing what a clinician should have done with the information available at the door.
And they raise a subtler problem: automation bias. When a system presents a priority score, humans tend to defer to it, even when they are supposed to be supervising. A model that “only decides with humans monitoring” can in practice become a model whose decisions are rarely overridden. The safety net assumed by supporters may not function the way it does on paper.
The middle positions most people actually hold
Few clinicians sit at either pole. Several intermediate positions deserve a vote of their own.
One is asymmetric authority: let an algorithm raise a patient’s priority on its own but never lower it without a human. This accepts more overtriage in exchange for protection against the most dangerous error, sending a very sick patient to the back of the queue. The cost is that overtriage already dominates human errors in the Kaiser data, and adding more can crowd the resuscitation areas it is meant to protect.
Another is restricted scope: allow automated triage for narrow, well-defined presentations with strong evidence, such as a clearly defined chest pain pathway, while leaving undifferentiated and complex cases to people.
A third is decision support with mandatory disagreement logging: the algorithm suggests, the nurse decides, and every disagreement is recorded and reviewed. Over time this produces the local evidence needed to judge whether the algorithm or the human was right more often, which is exactly the evidence both sides say is missing.
What regulators are already doing
The question is not purely hypothetical, and regulation is starting to shape the answer. In the United States, the FDA finalised guidance in December 2024 on predetermined change control plans, which let manufacturers of AI-enabled devices set out in advance how a model will be updated and validated after approval. That is an explicit acknowledgement that models drift and need maintaining. In the European Union, the AI Act treats AI systems that are medical devices, or safety components of them, as high-risk, bringing requirements for risk management, data governance, human oversight and post-market monitoring.
Neither framework answers the question in this debate directly. But both assume what opponents insist on: that an algorithm is not a finished product at approval, and that someone must keep checking it.
Second-order effects either way
Suppose AI triage is allowed to decide. Triage nurses, over time, may lose practice at the judgements they no longer make, which matters on the day the system goes down. Liability may push hospitals to defer to the algorithm even when staff disagree. And patients may learn to present their symptoms in ways the model rewards, as they already learn to do with human triage.
Suppose it is not allowed. The persistent errors of human triage continue, including differences by race, and it becomes harder to argue that a known, measurable failure is acceptable while a potentially better one is forbidden. Emergency departments short of experienced nurses may keep running with less expertise at the front door than an algorithm could offer.
Neither side can avoid costs. The debate is about which costs to choose and who bears them.
What would settle it
The most useful evidence would be a prospective trial in real emergency departments, comparing algorithm-led and nurse-led triage on outcomes that matter: time to critical intervention, deaths, and differences between patient groups. That trial would need to report results by race, age and sex, and to continue monitoring after it ends. Until that exists, both sides are arguing mostly from analogy.
Where the real disagreement lies
Put the two cases side by side and they are not as opposed as they first look. Both sides agree that current triage is imperfect. Both agree that algorithms can be biased and can fail. The real disagreement is about three questions.
- Which error is worse: the familiar, diffuse errors of human judgement or the less familiar, systematic errors of an algorithm?
- What standard of evidence should an AI triage system meet before it decides, rather than advises? Retrospective accuracy, prospective trials, or ongoing local monitoring?
- Who should be accountable when the decision is wrong?
Your turn to argue
We are not going to pick a winner. Tell us where you stand, and why.
If you are for AI triage: what evidence would you need to see before letting it decide? If you are against: what would change your mind, given how often human triage already gets it wrong? If you work in an emergency department, tell us what you have seen in practice. The best arguments will be the ones that take the other side seriously.
