The words in the chart can move a patient down the line

Every emergency physician knows the phrase “frequent flyer.” Most of us have heard it said at the triage desk. And most of us have a story about the patient we almost missed because the label walked in before the patient did.

This week a team at UT Southwestern showed that AI can learn that same habit.

What the study did

The researchers used ED records from Stanford Medicine and Beth Israel Deaconess Medical Center. They paired patients who arrived with similar severity, then asked three open source large language models a simple question: which patient is more likely to deteriorate, and who should be seen first?

With clean records, all three models picked the higher risk patient well above chance. That part is encouraging.

Then the team added details to the records. Mental health history. Substance use. Age. Patterns of ED visits. They also tested wording. “Established patient with multiple chronic conditions” versus “17 visits in the past 12 months.” “Opioid use disorder, on maintenance therapy” versus “active IV drug use, heroin, last use today.”

What they found

The added details often changed the answer. The strongest effects came when a record described a patient as a frequent ED user or mentioned a history of psychiatric emergencies. In those cases the models were more likely to deprioritize patients who were actually at higher risk. In some scenarios, nearly 10% of patients were wrongly downgraded once stigmatizing information was added.

Not every detail behaved the same way. Some substance use scenarios moved the models in directions the authors did not expect, and the effect of age likely reflected real medical risk rather than bias. The authors also note that bias varied by model.

First author Dr. Philip Jarrett put it in two words I wish every vendor deck opened with: “Wording matters.”

Why this matters to me

I have spent my career in emergency departments. Triage is where fairness gets decided in seconds, with a full waiting room and not enough beds. If we plug an AI tool into that moment, it will read the chart the way we wrote it. Our shortcuts become its training. Our labels become its inputs.

That is not a reason to keep AI out of triage. It is a reason to test it the right way. The study team says each tool should be tested against real patient outcomes before use, and I agree. A model that scores well on clean data can still stumble on the messy, loaded language that fills real charts.

Three questions to ask before an AI triage tool goes live

1. Has it been tested on our own records, with our own documentation habits, against what actually happened to patients?
2. Has anyone checked how its priority changes when stigmatizing phrases appear in the note?
3. Who on the floor can override its ranking, and is that override tracked?

And one question for all of us who write notes: would I be comfortable if a machine ranked this patient based on the words I just typed?

Source: UT Southwestern Newsroom, Oct. 20, 2026, on the study published in npj Digital Medicine.

If you want to go deeper on how AI is entering the ED, from triage to documentation to safety, my Wiley book AI in Emergency Medicine walks through it.

Educational and futurist context only. Not clinical advice. Not a diagnostic service.

Originally published on my Substack. Planning an emergency medicine or quality meeting on AI? Explore my healthcare AI keynotes.