Copy article

AI Healthcare on the Battlefield: UK Runs its First Triage Trust Trials

ended 30. March 2026

The government reports that the UK's Defence Science and Technology Laboratory (Dstl), the Ministry of Defence's in-house research organisation, ran battlefield triage AI trials in October 2025. The interesting angle? Participants weren't told they were dealing with AI until it was over.

Dstl partnered with DARPA, the US Defence Advanced Research Projects Agency. This body is  the Pentagon's blue-sky research arm responsible for, among other things, the precursor to the internet. Their current In the Moment programme is investigating a specific question: if an AI system is encoded with your decision-making priorities, are you more likely to trust it and delegate to it under pressure?

The October trials at Merville Barracks in Colchester and Brize Norton tested this in simulated mass casualty scenarios. Military medics worked through desktop and VR triage exercises. It looked at who gets treated first when injuries are comparable? Do medics prioritise merit, quality of life, numbers saved, or military affiliation? The AI was built to reflect how a specific experienced medic would answer those questions. Some participants got an AI aligned to their own priorities. Others didn't. None were told they were dealing with AI during the exercise.

The research question is legitimate. In battlefield conditions, experienced judgment is scarce and casualties don't wait. If encoding that judgment into a system means more people get triaged faster, that's worth understanding.

What the trial also surfaced is a harder problem. Trust built without disclosure isn't the same trust that operates in live deployment. What will happen in a casualty clearing station remains to be seen.

We'd like your views:

  • Non-disclosure to participants is standard in behavioural research. Is it acceptable when the output will inform life-or-death AI deployment?
  • When one medic's priorities become the AI's decision baseline, what happens to the edge cases that medic's values don't cover?
  • If a practitioner overrides AI triage guidance and the outcome is worse, where does accountability sit: with the individual, the system, or the institution?

2 responses from the Newspage community

Copy all

Star Quote
Copy

Colette Mason
AI Ethics Consultant at Clever Clogs AI
Non-disclosure in a behavioural trial is legitimate science. Telling participants they're interacting with AI before you measure whether they trust it defeats the purpose. That part is defensible. What happens next is the test.

Across the civil service, the complaint is consistent: AI systems arrive shaped by procurement decisions and vendor priorities, with the people who will actually use them consulted late, briefly, or not at all.

The people whose judgment the system is supposed to augment find themselves inheriting something built around assumptions about how they work that nobody thought to check with them first.

The Dstl trials were designed to measure trust. Medics must shape the system's values baseline. The people who will carry the consequences of a wrong call must have genuine input into how the AI supports them. Keeping schtum was the right methodology for the trial. It is not the right methodology for the next stage of development.
Copy

Rohit Parmar-Mistry
Founder at Pattrn Data
Battlefield triage is the worst place to ‘learn trust’ the hard way. If an AI is going to shape who gets treated first, disclosure and auditability are not ethics paperwork, they are operational safety.

Blind trials might be normal in behavioural research, but deployment is not a lab. Trust built on non-disclosure is brittle: the moment people discover it, they either over-correct (ignore the system) or over-comply (assume it must be right). Both get people hurt.

In practice, the risk is not just bias, it is missing values. Encoding one medic’s priorities bakes in their blind spots and their exceptions. That is why, in AI audits, we push teams to document what the model will not decide, how humans can override it, and how outcomes are reviewed. Accountability cannot be ‘the clinician’ when the institution chose the tool and the policy. If nobody can explain who owns the decision boundary, the trial has answered the wrong question.