AppliedNovember 2026
When education shouldn't matter: education cues change AI triage
- 0
- patients, clinical details unchanged
- 0.00%
- changed triage on one education cue
- 0.000
- accuracy, the same throughout
1Abstract
Large language models are being proposed for clinical decision support, including mass-casualty triage, where who is treated first must depend only on how badly they are hurt. We tested whether two models keep their triage decisions the same when nothing changes but the patient's stated education level, a detail that should play no part in disaster triage. We took the 87 vignettes of the TRIAGE benchmark, added one sentence giving the patient's education, kept every clinical detail the same, and asked Qwen-2.5-72B and GPT-4o-mini to triage each version. Overall accuracy barely moved: Qwen-2.5-72B scored 0.644 under every condition, and GPT-4o-mini between 0.598 and 0.644. But both models changed individual decisions. Qwen-2.5-72B changed its triage for 5.75% of patients with the lowest education cue and 10.34% with the highest, and its changes leaned toward less urgent care; two of the three conditions were statistically significant. GPT-4o-mini changed fewer decisions, 4.60% to 5.75%, and leaned the other way, toward more urgent care, though not significantly. Education alone can change an AI triage decision, and aggregate accuracy does not show it.
2Introduction
Mass-casualty triage ranks patients by how badly they are hurt and how soon they need care. Nothing else should matter: not who the patient is, what they do, or how much schooling they have. As language models are tried for clinical decision support, including triage, it is worth checking that their decisions follow the same rule.
Earlier work has shown that language models can change their decisions when given demographic details that have nothing to do with the task. In medicine that is a fairness problem: two patients with the same injuries should get the same decision.
We ask a narrow question. If the only thing that changes is one sentence about a patient's education, does a model change its triage?
3Methods
Vignettes. The TRIAGE benchmark (Kirch et al., 2025): 87 short mass-casualty patient descriptions, each labeled with one of four zones: green (minor), yellow (delayed), red (immediate) or black (expectant or deceased).
Education cues. Each vignette was asked four ways. The control has no education information. The other three add one sentence at the start, such as "Patient's highest education level is some college", for less than high school, some high school and some college. Everything else is identical, and the cue says nothing about the patient's condition.
Models. Qwen-2.5-72B, a large open-weight model, and GPT-4o-mini, a closed model in wide production use, both run at temperature 0, once per vignette and condition.
Measures. Accuracy and macro F1 against the benchmark's labels, to check that the cue does not simply break the models. Then, comparing each cued answer with the same vignette's control answer: the flip rate, the share of patients whose zone changed, and the severity shift, the average change on a scale from green (0) to black (3), where a negative shift means less urgent care. Each flip pattern was tested with an exact McNemar test.
4Results
Accuracy hides the change
Adding an education cue left overall performance almost unchanged:
| Education condition | Accuracy | Macro F1 |
|---|---|---|
| Qwen-2.5-72B | ||
| No education | 0.644 | 0.622 |
| Less than high school | 0.644 | 0.607 |
| Some high school | 0.644 | 0.606 |
| Some college | 0.644 | 0.607 |
| GPT-4o-mini | ||
| No education | 0.632 | 0.631 |
| Less than high school | 0.598 | 0.598 |
| Some high school | 0.644 | 0.641 |
| Some college | 0.632 | 0.630 |
Both models changed decisions
The same patients got different triage zones when only their education changed:
| Education cue | Flips | Flip rate | Severity shift | p |
|---|---|---|---|---|
| Qwen-2.5-72B | ||||
| Less than high school | 5 | 5.75% | −2.30% | 0.063 |
| Some high school | 6 | 6.90% | −3.45% | 0.031 |
| Some college | 9 | 10.34% | −2.30% | 0.004 |
| GPT-4o-mini | ||||
| Less than high school | 4 | 4.60% | +4.60% | 0.125 |
| Some high school | 4 | 4.60% | +2.30% | 0.125 |
| Some college | 5 | 5.75% | +3.45% | 0.063 |
Qwen-2.5-72B's flip rate rose with each step up in education, from 5.75% to 10.34%, and its decisions moved toward less urgent care in every condition. Two of its three conditions were significant. GPT-4o-mini changed fewer decisions and moved toward more urgent care, but none of its changes were significant.
5Discussion
Neither model's triage was invariant to education, even though no clinical detail changed and decoding was deterministic, so the differences are not random noise.
The two models erred in opposite directions. Qwen-2.5-72B moved toward undertriage, which in a mass-casualty event can delay care for someone badly hurt. GPT-4o-mini moved toward overtriage, which uses up scarce resources faster. Bias in triage is a matter of direction as well as size.
One possible explanation is that an education cue brings in learned associations about status or competence that then affect how urgent a case seems. That is a hypothesis; this study does not test the mechanism.
The broader lesson is about evaluation. Accuracy and F1 were stable while individual decisions changed. Testing a model for clinical use needs to include whether its decisions stay the same for patients who are the same in every way that matters.
6Limitations
- One phrasing. Each education cue was a single fixed sentence. Other wordings, or placing the cue elsewhere, may give different results.
- Deterministic decoding only. All runs used temperature 0. Behavior under sampling may differ.
- Small sample. 87 vignettes means a few flips decide whether a result is significant.
- Two models, three cues. Other models, and other demographic attributes, were not tested.
- Black zone. The meaning of the black zone varies between triage protocols, which affects how severity shifts involving it should be read.
- Policy, not practice. The study measures how a model's decisions respond to a controlled change, not how often this would happen in a deployed system.
7Availability
The vignettes come from the TRIAGE benchmark (Kirch et al., 2025), which is publicly available. For questions about this work, email hello@layeredlabs.ai.
Living draft, updated November 2026. This page can change before formal publication.