California bars AI from making clinical calls on its own -- then vetoes the bill that would have protected workers who override it

Governor Gavin Newsom signed two AI-in-healthcare bills on September 30 and vetoed two others the same day, and read together, the four documents do something California's AI lawmaking rarely does this plainly: they draw the actual line between what the state will mandate and what it thinks goes too far, in the governor's own words rather than a reporter's paraphrase.

AB 1979, effective January 1, 2027, requires any health facility, clinic, or physician's office to "take reasonable steps" to ensure a licensed professional "retains the ability to exercise independent professional judgment" whenever a patient's care is informed by a clinical decision support system -- AI that produces a prediction, recommendation, or analysis feeding into a diagnosis or treatment call. The bill bars facilities from letting any clinical decision rest solely on an AI tool's output, and separately designates health care chatbots marketed to consumers as "providers of health care" subject to the state's medical-confidentiality law. Routine, judgment-free automation -- appointment reminders, chart updates -- is explicitly exempted.

SB 503, also effective January 1, 2027, works the supply side of the same problem: it requires developers and deployers of a clinical decision support system to make "reasonable efforts" to identify risks of biased impact before the tool goes into a health program, publish a statement of the system's intended uses and known risks, and make reasonable efforts to mitigate the bias they find. Neither bill tells a hospital which AI tools to buy -- both tell it, and the vendor, what has to stay true no matter which tool it is.

SB 503's legislative path was fast for a bill carrying real compliance duties: the Assembly and Senate each passed it within a day of the other in late August, and it was enrolled August 27 -- about a month before the governor's signature. Its author, state Senator Akilah Weber Pierson (D-San Diego), is also a board-certified OB/GYN who practiced at Rady Children's Hospital before entering the legislature -- a detail worth noting given that the bill's bias-mitigation duty is written as a "reasonable efforts" standard rather than a pre-market certification requirement, the kind of calibration a clinician-legislator is positioned to make about what a hospital or developer can actually comply with.

The two vetoes are where this gets more interesting than a standard signing-statement roundup, because Newsom's own letters returning SB 903 and AB 2575 unsigned do not reject the bills' goals -- they specify, in detail, what each bill got wrong in the drafting. On SB 903, which would have prohibited companion chatbots from providing psychotherapy services and restricted licensed therapists' own use of AI, Newsom wrote that he supports "the author's intent to establish guardrails for the use of AI in healthcare settings," but that the bill was overly broad, would require routine AI-assisted screening determinations to get direct human approval, and defined "psychotherapy services" so loosely it would "capture general-purpose AI systems not even structured or deployed to deliver that care."

AB 2575 would have barred employers from retaliating against a direct-care worker for overriding a clinical decision support system's output, or for relying on one, when the worker's own judgment said that was necessary to meet the standard of care. Newsom's letter calls the underlying commitment -- that AI tools "support safe patient care and preserve professional standards" -- worth keeping, and the structural problem specific: an anti-retaliation rule that requires the state's Labor Commissioner to rule on whether an override met medical standards asks a labor regulator to practice medicine.

That third row is the gap worth watching. California now requires AI developers to hunt for bias in their own clinical tools and requires hospitals to keep a human in the final decision -- but a worker who actually exercises that judgment, overrides the machine, and gets punished for it by an employer has no AI-specific protection on the books, because the bill written to supply one was vetoed over how its enforcement mechanism was built, not over the idea itself. Newsom's letter explicitly invites "the author and stakeholders to revisit this issue next year."

None of the four bills addresses whether any currently deployed clinical decision support tool is actually, measurably biased -- that evidentiary question sits with whatever audits SB 503 now requires developers to produce starting next year, not with the legislature. And none of this week's action is California's first attempt at the problem: AB 1979 and SB 503 build on a state AI-lawmaking push Newsom's office has run through most of 2026, covering everything from workplace AI disclosure to deepfake labeling, now extended specifically into the exam room.