On the last day of the last legislative session of his governorship, the Governor converted California's health-care AI debate from a field of competing proposals into a legal architecture. Two health-care measures became law. Two companion measures, one protecting clinicians who override algorithms and one protecting the mental-health workforce, did not. What survived, and what was cut away, tells physicians more about the next decade than any single bill.
The record in one paragraph
On September 30, 2026, the Governor signed AB 1979 (Bonta) and SB 503 (Weber Pierson), both titled "Health care services: artificial intelligence," together with SB 947 (McNerney) on automated decision systems in employment and other AI-related bills (Governor's release). The same day he vetoed AB 2575 (Ortega), also titled "Health care services: artificial intelligence," and SB 903 (Padilla), "Mental health professionals: artificial intelligence" (legislative update). Three weeks earlier he had signed SB 813 (McNerney), creating a certification framework for independent AI verification organizations, and AB 1405 (Bauer-Kahan), creating a state registry of AI auditors. September 30 was the final day on which he could act on bills from the 2025–26 session.
What follows separates four layers that public coverage tends to compress into one phrase, "California AI law."
Layer one — the clinician's judgment and the chatbot's records (AB 1979)
Primary recordThe Governor's office describes the health-care pair as ensuring that physicians and other licensed providers can exercise their own professional judgment when AI or other clinical decision tools are used in patient care. AB 1979's sponsor, the California Nurses Association, describes the statute as barring health-care entities from using AI to replace the clinical judgment of licensed professionals (CNA statement).
Primary recordThe bill's second part reaches outside the hospital. As summarized during floor passage, a business that offers a health-care chatbot directly to consumers is treated as a provider of health care for purposes of the Confidentiality of Medical Information Act (Transparency Coalition). In plain terms, a consumer app that reads your medical records to answer your questions inherits the confidentiality duties that already bind your physician's office.
AnalysisThe second part may prove the more consequential. Clinical-judgment language restates what licensing law already presumes. Extending the medical-confidentiality statute to direct-to-consumer chatbots changes who is legally accountable for health data in a market that, until September 30, sat largely outside it.
Layer two — the developer's duty to look for bias (SB 503)
Primary recordSB 503 requires developers and deployers of AI-driven clinical decision support systems to make reasonable efforts to identify reasonably foreseeable risks of biased impacts arising from deployment in health programs (Transparency Coalition). The bill passed the Assembly 70–1 on August 24 and the Senate concurred 39–0 the following day (The Observer).
Analysis"Reasonable efforts" is a process standard, not an outcome standard. It does not prohibit a biased model; it obliges the developer and the deployer to have looked. Its practical force will depend on what documentation regulators and plaintiffs can demand.
Layer three — who checks the checkers (SB 813, AB 1405)
Primary recordSB 813 establishes a framework for certifying independent verification organizations to assess AI systems and models for safety and risk; AB 1405 establishes a state registry of AI auditors with standards for independence, transparency and integrity. On September 18, the Governor issued an executive order to accelerate implementation of both and convened an expert group to report within two months.
AnalysisSB 813's certification program is voluntary: it does not oblige any developer or deployer to undergo an audit. These are not health-care statutes, but they supply the machinery health-care AI has lacked: a recognized class of auditors whose findings a hospital, a health plan or a regulator could rely on. If SB 503's bias duty is the obligation, SB 813 and AB 1405 are the beginnings of a means of showing compliance.
Layer four — the workplace (SB 947, SB 951, AB 1331, AB 1883)
Primary recordSB 947 prohibits employers from relying solely on AI when making disciplinary or termination decisions. SB 951 requires notice when a mass layoff, relocation or termination is caused by an AI system. AB 1331 bars workplace surveillance tools in workplace bathrooms. AB 1883 limits workplace surveillance tools, including, as reported, employer predictions about a worker's emotional state and the collection of neural data.
ReportedCalMatters reports that SB 947 was narrowed during passage: it lost an appeals process for workers, the right to sue to compel compliance, and coverage of contractors (CalMatters).
AnalysisFor employed physicians, SB 947 matters more than its nickname suggests. Productivity dashboards, RVU-based performance flags and automated scheduling compliance are already part of employed practice. A statute requiring human review before discipline does not dictate the outcome of that review, but it creates a record that review occurred, or did not.
What the Governor rejected
AB 2575 (Ortega) — Vetoed. As analyzed by the Board of Behavioral Sciences in April, the bill would have barred developers and deployers from defending a claim on the ground that the AI acted autonomously, or that a clinician's failure to override the AI caused the harm (BBS analysis). It would also have protected health-care workers from retaliation for overriding AI outputs and required disclosure of AI use in patient care. The California Hospital Association urged a veto, arguing that the anti-retaliation provisions would prevent corrective action even where an override harmed a patient, and reported that the bill cleared the Legislature by a single vote (CHA). After the veto, the California Nurses Association said clinicians are left without explicit protection when they override unsafe AI (CNA).
SB 903 (Padilla) — Vetoed. CalMatters describes the bill as a measure to prohibit replacing mental-health workers with AI.
The Governor's veto messages for both bills are posted as image-only documents. Their stated reasons will be reported here once they have been transcribed and read in full.
The gap the architecture leaves
AnalysisRead together, the signed and vetoed bills draw a precise line. California law now says a licensed professional may exercise independent judgment when a clinical decision support tool informs care. It does not say the professional may not be disciplined for doing so, and it does not stop a developer from pointing to the clinician's failure to override as the cause of harm. The right to judge has been written into law; protection for judging against the machine has not.
That asymmetry will matter most where the clinician is employed and the algorithm belongs to the employer: staffing acuity scores, sepsis alerts, prior-authorization engines, discharge-timing tools. It is also where accountability will be contested first, in employment proceedings and malpractice litigation rather than in regulation.
What a physician or practice should do now
- Inventory the clinical decision support tools in use, who supplies them, and whether any of them reads patient records on a consumer's behalf.
- Ask vendors in writing what bias-identification process they follow and what documentation they will provide once SB 503 is operative.
- If you are employed, keep contemporaneous notes when you override an algorithmic recommendation on clinical grounds. AB 1979 recognizes your judgment; the vetoed AB 2575 would have protected you for exercising it.
- Do not assume any new duty binds before its operative date. Under California's default rule, these statutes take effect on January 1, 2027, unless the text provides otherwise.