ABSTRACT
A hospital deploys a software system that recommends treatments to clinicians. The system produces a recommendation that is clinically harmful but formally indistinguishable from a correct one. The clinician follows it. The patient is harmed. Who is liable?
Under existing legal doctrine, the answer is reliably the clinician. Developers invoke software-as-service classification. Deploying institutions point to contractual allocation. Regulators cite procedural conformity. The clinician, the only actor with a non-rescindible duty of care, absorbs what remains.
This paper argues that this outcome is not a distributional problem awaiting legal reform. It is a structural consequence of deploying systems whose architecture defeats the preconditions of accountability. Liability doctrines presuppose that adverse events are reproducible under investigation, that errors have reconstructible causal paths, that the failure space is enumerable, and that the system causing harm is the system that was authorized. Clinical large language models violate all four conditions simultaneously, and the violations are constitutive of the architecture, not contingent limitations.
The analysis traces how each major doctrine collapses on these properties, maps the evasion chain through which responsibility defaults to the clinician, and identifies configurations where liability remains viable. The prior question is whether systems whose architecture precludes reconstructible attribution should be authorized for high-stakes clinical delegation at all.
Natangelo, Stefano, Nobody’s Fault: Where Liability Collapses in Clinical LLM Deployment (May 23, 2026).
Leave a Reply