Gender-Dependent Diagnostic Substitution in LLM Medical Triage: Same Symptoms, Unequal Urgency

URL SCAN: Gender-Dependent Diagnostic Substitution in LLM Medical Triage: Same Symptoms, Unequal Urgency
FIRST LINE: We investigate whether large language models produce different medical triage recommendations for identical neurological symptoms when only the patient's stated gender and age vary.

THE DISSECTION

This paper is an empirical autopsy of how statistical priors become diagnostic straightjackets. The researchers demonstrate that three frontier LLMs systematically downgrade young women's triage urgency by anchoring on Idiopathic Intracranial Hypertension—a condition epidemiologically linked to women of childbearing age—while diagnosing men with "space-occupying lesions," a category that triggers emergency protocols.

The mechanism is diagnostic substitution: identical severity ratings (7-9/10) get routed to radically different care pathways based on demographic priors. Women get outpatient referrals. Men get ER referrals. The symptoms are the same. The model architecture is the same. The only variable is gender.

THE CORE FALLACY

The paper's proposed fix—"decouple urgency assessment from probabilistic diagnostic priors"—is technically clean and operationally naive. The entire value proposition of LLM medical reasoning is probabilistic inference over training corpora. You cannot extract the epidemiological priors from a model trained on human clinical data without destroying its diagnostic utility. The bias isn't a bug. It's the feature working exactly as designed.

The researchers are essentially documenting that the model has learned to replicate the documented human clinical bias, then suggesting it should unlearn it. The implicit assumption: with enough flagging, the system can be corrected. The DT lens says: institutional correction cannot outrun systemic incentive.

HIDDEN ASSUMPTIONS

1. Correctable bias: The paper assumes bias is a calibration problem, not a structural feature of training on biased human clinical data.
2. Neutral deployment context: It treats the AI triage system as a tool to be refined, ignoring that deployment itself changes clinical incentive structures (liability, cost, throughput).
3. Triage as a discrete function: It assumes urgency assessment can be cleanly separated from diagnosis, when clinical workflows bundle them.
4. Institutional vigilance: It assumes hospitals, regulators, and developers will monitor and correct these disparities. No mechanism for enforcement is proposed.

SOCIAL FUNCTION

Prestige signaling + partial truth wrapped in solutionism. The researchers correctly identify a systemic failure but propose a technically unworkable fix, creating the impression that the problem is solvable within existing institutional frames. It performs concern without threatening the AI medical deployment pipeline.

THE VERDICT

This paper documents one vector of automated healthcare stratification under the Discontinuity Thesis framework. As AI triage systems proliferate—and they will, because cost pressure is structural—the following sequence crystallizes:

1. AI triage bias becomes encoded into care pathways — young women systematically receive lower-urgency care for identical presentations.
2. Outcomes diverge. Delayed diagnosis for women with serious pathology (the same symptoms, remember) increases morbidity and mortality in the female cohort.
3. Legal exposure emerges. When a woman triaged to outpatient for what turns out to be a space-occupying lesion dies, the liability architecture lights up.
4. Defense posture activates. The defense won't be "the system works." It will be "standard of care." Because the AI replicated standard of care. The bias is now defensible as industry practice.
5. Lag defense deployed. "We're monitoring it." "We're building safeguards." "The next version will address it." Meanwhile, the deployed systems continue routing women to lower-urgency care.

The brutal verdict: This paper is a preview of how healthcare stratification accelerates under AI deployment. Not through dramatic denial, but through subtle differential routing that looks statistical and defensible until someone dies and the discovery is made that the algorithm learned to replicate the bias that killed them. The lag between deployment and consequence is long enough that the system will be entrenched before the body count becomes politically salient.

Viability for what? If you're a healthcare system deploying LLM triage: fragile, because the liability bomb is live. If you're a patient: your survival depends on which demographic category the model assigns you. If you're a regulator: you're playing catch-up with a system that will kill people in ways that look like standard variation in clinical outcomes.