AI Morbidity and Mortality: A Framework for Clinical AI Failure Review

URL SCAN: AI Morbidity and Mortality: A Framework for Clinical AI Failure Review
FIRST LINE: # Computer Science > Artificial Intelligence

The Dissection

This is an institutional damage-control protocol for deploying increasingly autonomous clinical AI without allowing individual failures to halt adoption. It converts messy AI-mediated harm into a reviewable chain: trigger, mechanism, clinical pathway, corrective action. The framework is operationally coherent, but its horizon is deliberately narrow. It asks how to repair failures inside the care system, not whether the system is being reorganized around machine authority and concentrated ownership.

The reported evidence is thin. Five illustrative cases and unanimous agreement across 20 classifications demonstrate that two reviewers can apply the rubric consistently. They do not demonstrate improved patient outcomes, causal validity, institutional learning, reduced recurrence, or generalizability across systems.

The Core Fallacy

The central error is governance substitution: treating better incident reconstruction as if it can convert a structural transformation into a manageable safety problem.

Under the Discontinuity Thesis, clinical AI failure is only the visible local hazard. The larger mechanism is the progressive transfer of cognitive work from human labor to AI capital. A stronger failure-review apparatus can reduce deployment risk, liability, and institutional resistance. It may therefore accelerate the very automation that destroys the mass employment-to-consumption circuit. The paper improves the machinery’s reliability while ignoring who owns it, who becomes economically unnecessary, and who retains control when human judgment is no longer the scarce input.

Its “blameless” posture is appropriate for learning from system failures, but it also launders power. Responsibility is redistributed across workflows, tools, and controls until the owners of the automating system disappear from the frame. The patient remains exposed; the institution becomes more deployable.

Hidden Assumptions

The abstract smuggles in several assumptions:

• That clinical AI remains a tool embedded in a fundamentally human-controlled workflow.
• That failures are discrete incidents rather than recurring expressions of broad model, data, incentive, and coordination problems.
• That institutions can preserve stable human oversight as AI systems become cheaper, faster, and more capable.
• That corrective actions will be implemented rather than documented as procedural theater.
• That the relevant objective is safer AI adoption, rather than the distribution of authority and productive necessity created by adoption.
• That prospective evaluation can validate the framework without confronting the political economy of automation.

The phrase “complement, rather than replace” is revealing. It limits the framework’s ambition to adding another layer of institutional control. It does not address the replacement of human cognitive labor—the discontinuity that matters.

Social Function

Classification: partial truth and transition management, wrapped in ideological anesthetic.

The framework is a legitimate local safety instrument. It can expose workflow defects and prevent some patients from becoming unrecorded casualties of automation. But its systemic function is to make AI deployment governable enough to continue. It gives institutions a vocabulary for absorbing failures, preserving legitimacy, and normalizing machine-mediated care. This is hospice care for the old human-centered workflow: competent, documented, and unable to reverse the underlying succession.

The Verdict

The paper diagnoses how AI can injure patients inside existing institutions, but not how AI reorganizes the institutions themselves. It treats the problem as failure containment when the strategic event is authority transfer and labor displacement. Useful as transition management; irrelevant to the terminal question of who controls clinical AI and what happens when human productive participation is no longer required.