Who answers when clinical AI causes harm? AB 2575 and the one defence it removes

California Assembly Bill 2575 was presented to the governor on September 15. If it becomes law, a party that developed, modified, selected or deployed a clinical decision support system will lose one defence: that the clinician’s failure to override its output broke the chain of responsibility and severed the defendant’s liability.

The bill is an argument about where risk settles. A system produces an output, a clinician acts on it or does not, and when the outcome is bad someone has to answer. Having a human in the loop can become an argument for placing responsibility there. AB 2575 treats that as a choice about where to put it, and a choice can be made differently.

AB 2575 is sponsored by the California Nurses Association; its companion, AB 1979, was last week’s Tuesday Standing Order. The protections are written from the point of view of the person at the bedside who has to decide whether to accept what the algorithm recommends or overrule it.

The bill adds Civil Code section 1714.48. Its reach is set by four verbs: developed, modified, selected, or deployed. The list reaches beyond the vendor to whoever changed the system, chose it or put it into use. Selected is the verb doing the new work in this argument, because choosing involves no engineering at all. A health system that picked this tool over another is named in the provision on the strength of that choice.

The definition covers AI outputs that aid clinical decisions about timing of care, diagnosis or treatment. It excludes specified administrative and educational activities when their independent performance does not require a professional licence.

It does not create a new right to sue. A patient still has to prove her claim. What changes is one move on the other side: treating the clinician’s failure to override as a superseding cause that cuts off the defendant’s liability. Other defences remain available, including evidence about causation, comparative fault, and reasonable precautions and disclosures. The bill does not make the clinician’s conduct irrelevant.

Subdivision (d) excludes malpractice actions against licensed or certified professionals providing direct patient care from this provision. Labor Code section 2821 separately protects workers from employer retaliation based solely on their override of, or reliance on, an output while exercising professional judgment within their scope of practice. The standard of care and other professional duties remain. A policy requiring deference to an output cannot supply the duty whose breach justifies discipline.

The distinction matters. The worker is protected in exercising judgment, not excused from exercising it. And a defendant can still contest responsibility without treating the human at the bedside as the break that severs its own.

The verb that should worry hospital administration is “selected”. Procurement involves contracts, security review, a signature. This bill explicitly names the choice of system, which should bring the record of why this tool and not another into sharper institutional focus. That record is not a new documentation requirement imposed by this bill. It is the question the bill should make an institution ask of itself.

So: which committee at your institution holds that record, and would it survive being read aloud?


Sources. AB 2575 (2025–2026), enrolled text · Bill history


The Standing Order

Published Tuesday, Wednesday and Thursday on nonalgorithmic.com, and nowhere else. The essays arrive each Friday.

Leave a Reply

Discover more from Nonalgorithmic

Subscribe now to keep reading and get access to the full archive.

Continue reading