The Army's Best Medical AI Program Is a Bet Against the Department's AI Strategy

The most defensible thing anyone is currently doing with AI in military medicine works because it buys back time for people to practice without AI. The Department's own AI strategy is pushing in the other direction as hard as it can. Both of those are documented. Neither document mentions the other.

Two facts that belong in the same document

One. The Joint Special Operations Medical Training Center at Fort Bragg runs the nearly nine-month Special Operations Combat Medic pipeline and the four-month Special Forces Medical Sergeant Course. Its instructors cannot make the calendar longer, so they are using analytics to decide what to cut. Leaders enter granular student performance data — time to master each skill, repetitions required — into what the course chief calls a calculator, and the aggregate across course iterations tells them where the schedule is being wasted. Business Insider reported it in its July 2026 piece on the school's use of AI-assisted analytics: the goal is not to replace instructors but to find hours. The course chief, who spoke on condition of anonymity, put it this way: "We can shift resources and time to put the emphasis where the students need it and where their areas of weakness are." The worked example the school gives is tourniquets. Students tend to arrive already familiar with application, or to master it faster than other skills, a legacy of two decades of war in which application was effectively the whole skill — and the time that frees can go instead toward tourniquet removal, along with infection and medication management. There is no model in this, no inference, and no clinical risk. The output is a schedule.

Two. Krzysztof Budzyń and colleagues published what Douglas Rex describes as the first large-scale formal evaluation of whether AI causes physician deskilling. In their multicentre observational study in The Lancet Gastroenterology & Hepatology, adenoma detection rate for colonoscopy performed without AI fell across four Polish centres from 28.4 percent before those centres adopted AI polyp detection to 22.4 percent after — an absolute difference of −6.0 percent (95% CI −10.5 to −1.6, p=0.0089), odds ratio 0.69 in multivariable analysis. Fifteen of the nineteen endoscopists showed numerical decreases. Rex calls the results very concerning. These were not trainees: each had performed more than two thousand colonoscopies, with a mean of twenty-eight years of experience.

Put plainly: the Army is spending reclaimed hours on unassisted judgment. Budzyń is the best-known evidence that continuous exposure to assistive AI is associated with that judgment decaying, in experienced specialists, on the task they were most practiced at. It is one observational result in colonoscopy. A later prospective trial did not reproduce a skill drop on a coarser endpoint, and neither paper is a conversion decision.

The schedule at Fort Bragg does not need that dispute resolved. I think it is the single most defensible thing anyone in this field is currently doing with AI, and it is defensible for a reason that has nothing to do with the analytics being clever.

Why tourniquet conversion is the right test case

Discipline does not tell you where to set a decision threshold. That is the whole problem with conversion, and it is why conversion — rather than triage scoring, or imaging, or a chatbot — is the case that should decide how fast military medical AI gets fielded.

Conversion is the headline change in the current doctrine. The TCCC Guidelines of 1 May 2026, summarized by Deaton, Montgomery and Butler in the Journal of Special Operations Medicine, moved tourniquet reassessment from an afterthought into a governed procedure with time windows and preconditions. Deployed Medicine is the authoritative platform for the text. The operational driver is evacuation. Business Insider's reporting on how the war in Ukraine is changing U.S. tourniquet training describes a battlefield where drones have made timely casualty evacuation unreliable, and where a device applied correctly and left in place becomes its own injury — rhabdomyolysis, dialysis demand, nerve damage, limb loss. John Holcomb, formerly head of the Army's Institute of Surgical Research, gives the number: "When you don't control the air, when you have denied, prolonged evacuation, then you put a tourniquet on, if you leave it on for longer than two hours, it can cause more problems than the wound." The same reporting notes that the next war may leave wounded troops waiting weeks for evacuation, which is the planning assumption underneath all of it.

So the skill is now graded. Students in the pipeline are assessed on removals and on their ability to teach removal to others. Col. Ken Dwyer, who commands the Army's Special Warfare Center and School, describes the decision the student has to own: "I put on a tourniquet to save this guy's life. But at some point, I need to figure out a less destructive way to control that bleeding. I have to remove that tourniquet and see if a pressure dressing will work or something else will work, and to try to save that limb." Stacy Shackelford, a professor of emergency medicine and surgery and the former director of the military's Joint Trauma System, calls the shift a 180: "Every single person who's trained to put on a tourniquet needs to also be trained to take off a tourniquet."

Now notice what kind of decision that is. Application is a motor skill with a binary success criterion — the bleeding stops or it does not. Conversion is a threshold call made under time pressure with incomplete information, where both errors are expensive and the costs are not symmetric. Loosen too early and the casualty bleeds. Leave it and the limb goes. The inputs are the wound, the patient's condition, what you can monitor, and how long you think you are going to be holding this person. None of those arrive as numbers.

This is precisely where an app is least useful and most attractive. Shackelford makes the distinction directly. Reporting on the AI tools Ukrainian and U.S. teams are building for troops treating wounds alone, Business Insider records her point that an app can plausibly help a user remember when a tourniquet went on, but that deciding whether it can safely come off requires judgment about the wound and patient condition that is difficult to teach in basic first-aid training, let alone through an app. She is not against the tools. She describes them as "decision assist," a way to augment memory or help with difficult choices, and observes that the realistic alternative for an untrained service member is being told, "Okay, you're on your own."

Both halves of that are true, and the gap between them is where the timer lives. An app that tracks application time is a memory aid and almost certainly a net good. An app that recommends conversion is a threshold-setting machine pointed at the one skill the school just spent its reclaimed hours trying to build.

Memorizing the sequence makes you a technician. Knowing why it is ordered that way is what lets you recognize the specific case where the sequence is wrong and deviate on purpose. That is what conversion training is buying. Whether an assistive tool takes it back is an open measurement, not a result this literature has settled.

What the reclaimed training hours used to buy

I qualified as a Special Forces Medical Sergeant (18D) through the Fort Bragg pipeline and served with 7th Special Forces Group (Airborne). I later developed and delivered tactical medical training for Special Forces medics across six multinational units, and completed two deployments to Colombia and two deployments to Afghanistan as a U.S. Army Special Forces Medical Sergeant (18D). So I have sat on both sides of the schedule the analytics are now optimizing, and what that warrants is an observation about cost rather than about me. Application reps are cheap to teach and cheap to grade: you can run many of them per instructor-hour, and the assessment is objective. Judgment cases are the opposite. To teach conversion you have to generate varied presentations, let students commit to a call, and then adjudicate the disagreements, including the ones between instructors. That is expensive, it scales badly, and it is the first thing that gets compressed when a calendar is full. The reason the reclaimed hours matter is not that there are more of them. It is that they are being moved from the cheap category into the expensive one, which is the direction training schedules almost never move on their own.

Where a 30-day model refresh collides with FDA validation

Set the training picture aside and look at what the Department is telling vendors to do.

The Artificial Intelligence Strategy for the Department of War, issued 9 January 2026, is unusually specific about cadence. Under the heading of AI model parity it directs the Chief Digital and Artificial Intelligence Office to "establish a delivery and integration cadence with AI vendors that enables the latest models to be deployed within 30 days of public release," and states flatly: "This shall be a primary procurement criterion for future model acquisition." Model-objectivity benchmarks become a procurement criterion within 90 days, and "any lawful use" contract language follows within 180. Nextgov's analysis of the acceleration strategy reads the 30-day clause as the sharpest break in the document, against a baseline in which fielding a frontier model inside a government enclave has historically taken more than a year.

The FDA's regime for software as a medical device wants the opposite thing: a model that is frozen, validated, and changed only in ways specified in advance. The mechanism built to reconcile those two postures is the Predetermined Change Control Plan, and it exists — the agency's final guidance on marketing submission recommendations for a PCCP for AI-enabled device software functions has been out since August 2025. A PCCP requires a description of the planned modifications, a protocol for developing and validating them, and an impact assessment. It is a way to pre-authorize a bounded set of changes. It is not a license for continuous learning, and it certainly is not a thirty-day refresh against whatever a lab shipped last week.

Then check whether anyone in this domain is using it.

APPRAISE-HRI is the military hemorrhage-risk software people usually have in mind when they say battlefield AI is already fielded. Its FDA 510(k) record, K233249, decided 5 April 2024, lists the applicant as the Surgeon General of the Department of the Army. SpectralMD's DeepView, authorized this May, holds De Novo DEN250028 for adjunctive burn wound assessment. Both records carry the same line: Predetermined Change Control Plan Authorized: No.

The mechanism exists. In military trauma AI, it is not being used. Nobody has demonstrated the paperwork path from a 30-day model refresh to a lawfully marketed medical device, and a procurement criterion does not create one.

The 30-day cadence is already real somewhere else. On 17 August 2026 the Chief Digital and Artificial Intelligence Office said it was deploying Gemini 3.7 Flash on the commercial launch day. That is workforce chat. APPRAISE-HRI and DeepView still have no authorized Predetermined Change Control Plan. Enterprise tools are on the procurement clock. Trauma devices are not.

Two more facts belong here because they set the evidentiary floor. The Congressional Research Service's June 2026 brief on FDA regulation of AI-enabled devices states that the agency "does not appear to have authorized for marketing any generative AI (genAI)-enabled device" — so a language-model conversion advisor sits in a category with zero authorized precedents. On 18 August 2026 the FDA published Considerations for the Regulation of Generative AI-Enabled Medical Devices, a discussion paper under docket FDA-2026-N-7874, with comments due 19 October 2026. It is not guidance and it authorizes nothing. It is the document in which the agency asks how a language-model device, a foundation-model update, and an agentic system would be evaluated. And a systematic analysis in PLOS Digital Health covering all 1,357 FDA-cleared AI/ML devices through 5 December 2025 found that only 3 of them — 0.2 percent — had been evaluated for patient-centered outcomes such as mortality, morbidity, or readmissions. Thirty-four were linked to registered prospective trials; twelve posted results. The authors' conclusion is the one to keep: readiness should no longer be defined by FDA clearance alone.

A treatment that works in a study and cannot be delivered in the field is not a treatment. The regulatory path is the delivery constraint here, and a model that satisfies a benchmark but cannot be lawfully fielded at the cadence its own acquisition policy demands has the same problem.

What I would fund, and what I am predicting

If I had the money, I would spend it in this order.

The analytics, because they are already working. Learning analytics with no inference in the clinical loop, producing a reallocated schedule, is the rare military AI program with a measurable output and no patient exposure. It should be instrumented properly: publish the reallocation, and track whether graduates' unassisted conversion decisions actually improve. Right now the reallocation is the claim and the outcome is unmeasured.

The deskilling study nobody has run. Budzyń et al. is an observational study in gastroenterology, susceptible to confounding. A September 2025 correction restored an omitted covariate, indication for colonoscopy, and the authors reported that the interpretation did not change. Pedersen, Mori, and colleagues then ran a pragmatic prospective trial: after computer-aided detection was removed, unaided detection of polyps at least 5 millimeters showed no deskilling and no upskilling. Inexperienced endoscopists detected more only while the tool was on. Experienced endoscopists, closer to Budzyń's cohort, showed no significant change in that rate. A null on that polyp endpoint does not erase an adenoma-detection drop, and neither paper is a conversion decision. What the pair does is retire the claim that decay is uncontested. The military version of the question is still unrun, and it is cheap by defense standards: take two cohorts through conversion training, one with a decision-assist app and one without, and test both unassisted on cases the app has not seen.

Memory aids, not recommenders. Goddard, Roudsari and Wyatt's systematic review in JAMIA screened 13,821 papers, included 74, and found that erroneous clinical decision support increased the risk of an incorrect decision by 26 percent (risk ratio 1.26, 95% CI 1.11 to 1.44). The useful part is the mitigator list, which reads as a design specification: emphasize user accountability, position advice carefully, attach current confidence to output, and present supporting information rather than a recommendation. Lyell and Coiera add that automation bias appears in single tasks rather than only under multitasking, shows up more in diagnosis than in monitoring, and rises with the difficulty of verifying what the automation said. A tool that says "this tourniquet has been on for 2 hours 14 minutes, here is what the guideline conditions on" satisfies that specification. A tool that says "convert now" violates every item on it.

I would rather ship a narrower system that refuses than a broad one that guesses confidently.

That is a preference, and preferences are not arguments. So here is the part that can be wrong.

Prediction one. If anyone runs the cohort study above, the app-trained cohort will score better on recall of the protocol's numeric thresholds and worse on unassisted conversion decisions against novel cases. The Pedersen trial failed to find a skill drop on polyp detection after the tool came off. That is a different task, and it does not retire this prediction. Falsified by a trial in which the app-trained cohort performs as well as or better than the control on unassisted conversion quality.

Prediction two. On 1 January 2029, no AI-enabled medical device whose FDA indications for use name prehospital, point-of-injury, or battlefield trauma triage will hold an authorized Predetermined Change Control Plan. Enumerate the candidates from the FDA's list of AI-enabled medical devices; falsified by one whose clearance record reads "Predetermined Change Control Plan Authorized: Yes."

If prediction two holds, then the 30-day model parity criterion and military medical AI will have spent three years in the same strategy document without ever meeting, and the Department will have kept fielding medical software the slow way while describing itself as AI-first. If prediction one holds, the schedule at Fort Bragg is worth more than any triage model currently in development, and the correct move is to buy more of it.

I have written at length about what the field actually looks like underneath this argument — what is cleared, what is experimental, and what the triage evaluations show — in AI-driven medical triage and diagnostics, and about edge deployment constraints in AI at the operator's edge. Background is on the about page.

References

1. Baker K. "Modern war demands more of special operations medics. AI is helping the Army cram in more of what they need most in training." Business Insider, 24 July 2026.
2. Baker K. "The war in Ukraine is exposing the dangers of prolonged tourniquet use. US special operations medics are adapting." Business Insider, 19 July 2026.
3. Baker K. "Ukrainian and US developers are building AI tools for wounded troops when there is no medic nearby." Business Insider, 7 September 2026.
4. "The next war may leave wounded troops waiting weeks for evacuation." Business Insider, August 2026.
5. Budzyń K, Romańczyk M, Kitala D, et al. "Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study." Lancet Gastroenterol Hepatol 2025;10(10):896–903.
6. Rex DK. "Does exposure to CADe decrease detection skills without CADe?" ASGE Journal Scan, reviewing Budzyń et al.
7. Deaton TG, Montgomery HR, Butler FK. "Tactical Combat Casualty Care (TCCC) Guidelines: 1 May 2026 Updates." J Spec Oper Med 26(2):89.
8. Deployed Medicine — canonical TCCC platform (login required).
9. Artificial Intelligence Strategy for the Department of War, 9 January 2026 (PDF).
10. "DOD's AI acceleration strategy." Nextgov/FCW, February 2026.
11. FDA. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions, final guidance, August 2025.
12. FDA. 510(k) K233249 — APPRAISE-HRI, decision 5 April 2024.
13. FDA. De Novo DEN250028 — DeepView AI System, granted 21 May 2026.
14. FDA. Artificial Intelligence-Enabled Medical Devices list.
15. Congressional Research Service. IF13245, FDA Regulation of AI-Enabled Devices, 10 June 2026.
16. Abulibdeh R, Cajas Ordóñez SA, Celi LA, et al. "1,357 AI medical devices cleared, 3 actually tested on patient outcomes." PLOS Digit Health 2026;5(8):e0001597.
17. Goddard K, Roudsari A, Wyatt JC. "Automation bias: a systematic review of frequency, effect mediators, and mitigators." JAMIA 2012;19(1):121–127.
18. Lyell D, Coiera E. "Automation bias and verification complexity: a systematic review." JAMIA 2017;24(2):423–431.
19. Correction to Lancet Gastroenterol Hepatol 2025; 10: 896–903. Lancet Gastroenterol Hepatol, 11 September 2025.
20. Pedersen TA, Mori Y, Botteri E, et al. "Learning and deskilling effects of artificial intelligence in colonoscopy among endoscopists with different levels of experience: a pragmatic, prospective trial." Endoscopy 2026.
21. FDA. Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback, 18 August 2026. Docket FDA-2026-N-7874; comments due 19 October 2026.
22. Department of War, Chief Digital and Artificial Intelligence Office. "Deploying Gemini 3.7 Flash in lock step with its commercial launch." 17 August 2026.

Views expressed here are my own and do not reflect official positions of the U.S. Army, the Department of Defense, or the U.S. Government. Nothing on this page is medical advice or clinical guidance. Protocol references are descriptions of published doctrine written for people who build software; anyone treating a casualty should work from the current official guideline and within their own scope of practice.