AI in Mass Casualty Triage: Promise, Peril, and Override

The drill ran on a Saturday morning. Forty simulated casualties. Eight EMS crews. A new AI-assisted electronic triage system the regional disaster coalition had asked us to evaluate. The system was fast. It tagged most of the immediate-category patients correctly inside ninety seconds of the initial sensor read. It also assigned a delayed tag to a young woman with an evolving tension pneumothorax because her oxygen saturation was still 94% on the first read. The medic who caught it overrode the algorithm and re-tagged her immediate. Two minutes later, on the second sweep, her saturation had dropped to 78%. The algorithm would have updated. The medic was already three patients ahead of it.

That is the entire conversation about AI in mass casualty incident triage in 2026. The systems are fast. They are also brittle. The question for emergency physicians, EMS medical directors, and disaster planners is not whether to use them. The question is what your department, your medics, and your protocols need to have in place before the first real incident, so that when the algorithm is wrong, the human catches it before the patient pays for it.

After running and reviewing AI-assisted MCI exercises with regional disaster coalitions over the last eighteen months, I want to walk through three things: what the 2025–2026 evidence actually supports, the override problem that is now formally recognized in the literature, and the operational framework I would put in writing before any AI-augmented triage system is fielded in your jurisdiction.

What the 2025-2026 Evidence Supports

The data is genuinely encouraging on the throughput side. A November 2024 BMC Public Health systematic review synthesized nineteen high-quality studies on AI in disaster and MCI triage and found consistent improvements in three areas: time-to-tag, completeness of vital sign capture, and continuous re-prioritization based on physiologic drift. Electronic triage tags that combine wearable sensors with embedded decision algorithms triaged roughly three times the volume of casualties per unit time compared to paper-based START triage in simulation, and they captured continuous vital sign data that paper triage simply cannot.

The deep learning side is also maturing. Several 2025 studies built deep neural network models for primary MCI triage that outperformed the Revised Trauma Score on AUC, with the best models reaching 0.89 against expert adjudication. A 2026 March arXiv benchmark paper introduced an open MCI-like field-regime dataset derived from MIMIC-IV-ED, restricted to the data actually available in a field environment — vitals, observations, brief notes — and showed that even with that limited input set, transformer-based models could meaningfully predict early deterioration. That benchmark matters because for the first time, MCI triage AI has a reproducible academic standard against which vendors can be measured. Demand it from your vendor by name.

The autonomous systems work is the most operationally significant 2025 development. DARPA's Triage Challenge concluded its second year in October 2025 at the Guardian Centers in Perry, Georgia, with unmanned aerial and ground vehicles running real-time casualty identification and injury assessment in scenarios modeled on a C-130 crash and night ambush. DART and MSAI took top honors in the Systems and Data competitions, respectively. Final competition is in 2026. The translation pathway from defense MCI triage to civilian disaster response is now formal, funded, and short.

That being said, the same evidence base is clear about the ceiling. The studies that show the best AI MCI performance are simulation studies, controlled exercises, or retrospective dataset reconstructions. Real-world MCI deployment data — actual incidents, with actual chaos, with actual triage tag use — remains thin. A 2025 Crisis Medicine survey of 464 EMS employees found that triage tags of any kind were used in 92% of MCI drills but only 34% of actual MCIs. That gap is the operational problem AI cannot solve by itself.

The Override Problem the Literature Now Names

The most important finding in the 2025–2026 MCI AI literature is not about the algorithms. It is about the humans operating them.

The override problem is now recognized as a distinct failure mode in AI-augmented MCI triage. Algorithms are excellent at the static read — what the patient looks like in this two-minute window. They are less good at the trajectory. A patient who is currently green-tagged but on a steep physiologic slope toward red is exactly the patient who kills you in an MCI, and the algorithm reading frozen vitals at minute three will not flag them until minute six. By minute six, three more casualties have arrived and the medic has moved on.

The Crisis Medicine analysis I cited above goes further. The article reviews a body of evidence that experienced EMS providers in real MCIs frequently rely on intuition-based triage that does not map cleanly onto formal protocols, and that this intuition often outperforms strict adherence to the printed algorithm in chaotic field conditions. The 2025 Kennesaw State human-AI collaboration study addressed the same question from the training side: when participants triaged ten simulated casualties with an active AI partner versus with no AI partner, the collaboration group performed better only when participants felt comfortable overriding the AI on difficult cases. Reliance on AI predicted performance only up to the point where the AI was right. Past that point, automation dependency degraded outcomes.

The implication is operational and uncomfortable. AI in MCI triage performs best in the hands of operators who have been trained, drilled, and given written permission to override it. The 2025 ethics framework for AI in emergency triage that I have been citing for civilian ED work applies here with double weight. Algorithm dependency is a recognized failure mode and it shows up fast under MCI cognitive load. For a longer treatment of how this plays out in mass casualty conditions specifically, see AI in Mass Casualty Triage: Promise, Peril, and the Human Override Problem on Medium.

An Operational Framework Before You Field It

If your jurisdiction is being approached by a vendor offering AI-assisted MCI triage hardware or software, here is what I would put in writing before any deployment.

First, the protocol owner is the physician medical director, not the vendor and not the IT department. Every AI MCI triage tool gets approved into the regional disaster plan only if the medical director signs off on the operational protocol around it — including the override pathway, the documentation standard, and the failover plan when the system is down. Vendors will not write this. You write it.

Second, override is one button or zero buttons, never two. Every medic in the field must be able to re-tag a patient with a single physical action that the system records as a deliberate override, with no friction and no documentation prompt at the moment of override. Documentation comes later, on the after-action report. Friction in the override path during an MCI is the single most dangerous design choice a vendor can make. Test for it before you sign.

Third, the AI tag is a starting point and the medic's tag is the operative tag. Document both. The AI tag is data. The medic tag is the standard of care. When they diverge, the medic decision controls and the AI tag is in the chart for after-action review. This protects the patient, the medic, and the system, and it generates the only data you will ever have to actually evaluate the algorithm against ground truth.

Fourth, drill it like you mean it. Quarterly MCI drills with the AI system active, scoring both the system's tag rate and the override rate. Track which kinds of cases the system misses. Feed the override patterns back to the vendor as a contractual deliverable. If your vendor is not interested in seeing your override data, that is your answer. The training side of this work is exactly what scenario-based simulation platforms are built for, and a structured curriculum is far more effective than ad-hoc tabletop drills. The EMS simulation environment at emsmedsim.globalmedopscommand.com is one of the platforms that supports this kind of repeated AI-augmented MCI drilling at scale.

Fifth, plan for the system being offline. The single most likely state of your AI MCI triage system on the day a real incident happens is unavailable, because the cell network is saturated, the power is out, or the field hub got knocked over by a fire engine. Paper triage tags, START algorithm cards, and a standard MCI triage worksheet need to be in every field bag and every command vehicle, and your medics need to be drilled on them at least as often as they are drilled on the AI system. The AI is an augment, never a dependency.

What You Should Be Doing Now

Audit your regional disaster plan for whether AI-assisted MCI triage is mentioned at all. If it is, look for whether the override protocol, documentation standard, and failover plan are written down in the plan itself, or whether they are deferred to the vendor. If the latter, that is your gap. Tabletop the worst-case override scenario at your next disaster coalition meeting — a green-tagged patient who deteriorates between sweeps and the algorithm did not update — and walk through who catches it, when, and how it gets re-tagged. If your jurisdiction has not yet adopted an AI MCI tool, this is the moment to set the procurement bar high. Demand published subgroup performance, demand the override-rate data from prior deployments, and demand a contractual commitment to share field override patterns back to the regional medical director.

Dr. Chet's Take

Mass casualty incident response is the single most chaotic operational environment a physician will ever be asked to lead. Every senior leader I trust who has run a real MCI says the same thing: the protocol is the floor, not the ceiling. Field commanders in critical care transport, air medical, and military disaster response have known for decades that you brief the protocol, you drill the protocol, and then you train the team to recognize the moment the protocol stops fitting the situation in front of them. AI does not change that calculus. It raises the floor. It does not raise the ceiling.

What I tell every EMS medical director and every disaster coalition chair I work with is this. Buy the AI system if it earns its way into your plan on documented field performance, override transparency, and vendor accountability. Do not buy it on a demo. Drill it like a real operational asset, not a piece of vendor demo gear. And tell your medics in plain language, in writing, that they are expected to override it the moment their hands tell them something the algorithm does not see. The patients you save in an MCI are the ones the medic catches between sweeps. That has been true since paper triage tags. It will be true after AI replaces them.

AI Won't Wait. Neither Should You.

If you are evaluating an AI MCI triage tool right now without a written override protocol, a documentation standard for divergence between AI and medic tags, and a quarterly drill cadence, you are exposed. Consider enrolling in my course: AI in Emergency Medicine: Becoming AI Bulletproof. The course walks through the exact override-protocol language we use in our regional disaster plan, the AI-versus-medic documentation standard that protects both the patient and the provider, the procurement checklist for AI MCI triage vendors, and the drill template I run with our coalition every quarter.

Learn more: AI in Emergency Medicine: Becoming AI Bulletproof.

If you're an emergency physician (or any clinician treating patients daily) trying to understand how AI will actually impact your clinical practice — not just the hype — I put together a free practical guide. You can download it here: AI in EM Survival Guide.

Dr. Chester "Chet" Shermer, MD, FACEP is a Professor of Emergency Medicine, TeleHealth, HEMS and Critical Care Transport, and State Surgeon for the Army National Guard. He is the founder of Global MedOps Command and the creator of AI in Emergency Medicine: Becoming AI Bulletproof.

Also on Medium: Read more from Dr. Shermer on Medium

Sources

1. BMC Public Health, "Application of artificial intelligence in triage in emergencies and disasters: a systematic review," November 2024, https://pmc.ncbi.nlm.nih.gov/articles/PMC11575424/
2. DARPA, "DART and MSAI triumph at DARPA Triage Challenge," November 2025, https://www.darpa.mil/news/2025/dart-msai-triumph-darpa-triage-challenge
3. arXiv, "LLM-Assisted Emergency Triage Benchmark," March 2026, https://arxiv.org/html/2509.26351v2
4. Crisis Medicine, "Rethinking Triage: Could Intuition Outperform Formal Triage Systems," September 2025, https://www.crisis-medicine.com/rethinking-triage-systems/
5. Kennesaw State University, "Enhancing Mass Casualty Triage Training Through Human-AI Collaboration," December 2025, https://digitalcommons.kennesaw.edu/undergradsymposiumksu/fall2025/fall2025/72/