Some of the ways AI fails are genuinely new. What makes them dangerous is not that they are exotic, but that they tend to surface as harms we already recognise. A missed result. A wrong dose. A late escalation. So the question worth sitting with is this. If a model now shapes that decision, who is governing the model?

The paper and its author

That question runs right through the paper, which works from the WHO's patient safety categories to show that AI does not stand apart from the familiar mechanisms of harm. It amplifies them.

The failure modes she names

Agnès's research identifies the failure modes plainly: workflow mismatch, automation bias, silent model updates, interoperability failures, data quality dependencies, system coupling. Each is new, and each amplifies a mechanism of harm the system already knows. A warning placed where no one looks. A result hidden behind a broken interface. A model updated quietly by its supplier, so the safety case signed at procurement no longer describes the system in use.

The silent failure loop How a clinical AI model harms a patient when no one owns the call Familiar harm, new route 1 Supplier updatesthe model quietly 2 Safety checksno longer match 3 Familiar harmslips through unseen 4 Filed as a glitch,nothing is learned STRASYS Our read of one failure mode in Leotsakos (2026): silent model updates.
An illustration of one failure mode the paper names: silent model updates.

Two implications for leaders

Two implications land hardest for leaders. AI-generated harm rarely comes from the algorithm alone; it comes from people, design, integration and supplier-controlled change. And our rules were written for static devices, not models that keep learning after they go live, which is how a tool reaches the ward with no post-deployment monitoring and no clear owner when it drifts. The paper's prescription is the line we would underline: AI safety has to be built into the patient-safety system an organisation already runs, not bolted on beside it. Said the way we say it at STRASYS, AI supports judgement, it does not replace responsibility.

Read the original

The paper goes further than a short note can. It maps each failure mode onto a specific category of patient harm, works through clinical examples, and sets out where governance needs to move next. If you are accountable for AI in a clinical setting, it is worth reading in full.

Read the full article on HealthManagement.org

"AI in Healthcare: Benefits, New Failure Modes and Implications for Patient Safety," Dr Agnès Leotsakos, HealthManagement.org The Journal, Volume 26, Issue 3 (2026). The rights belong to HealthManagement.org. Please read and cite the original.

Questions this raises

Workflow mismatch, automation bias, silent model updates, interoperability failures, data quality dependencies, and system coupling. Each is new to AI, but each amplifies a mechanism of patient harm the system already knows.
Because the safety case signed at procurement no longer describes the system in use. A supplier updates the model quietly, the safety checks written for the original version no longer match, and familiar harm slips through unseen.
AI-generated harm rarely comes from the algorithm alone. It comes from people, design, integration and supplier-controlled change. Boards need to build AI safety into the patient-safety system the organisation already runs, not bolt it on beside it.
AI supports judgement, it does not replace responsibility. Our Decision Intelligence Platform uses AI to surface patterns and evidence for leaders, but the decisions remain with the people accountable for them.