California AB 2575

Overview

California AB 2575 was vetoed by the Governor on 2026-09-30, so it never became law and none of its provisions apply. The enrolled bill would have addressed the use of artificial intelligence in health care services by creating two separate legal tracks: a Civil Code rule limiting certain liability defenses in actions involving clinical decision support systems, and a Labor Code worker-rights article protecting direct patient care workers from retaliation or discrimination when they exercise professional judgment rather than follow an AI output. The measure was introduced by Assembly Member Ortega on 2026-02-20 and, according to the enrolled text, was amended repeatedly before passing both chambers on 2026-08-31. Because the bill was vetoed, the operative legal effect is zero despite the detailed framework the Legislature approved.

The bill’s core policy choice was to preserve human professional judgment in clinical settings. It would have declared that workers providing direct patient care are free to make assessments and decisions within their scope of practice as appropriate for their patients, and it would have prohibited an employer from retaliating or discriminating against a worker based solely on the worker’s override of, or reliance on, the output of a clinical decision support system. In parallel, it would have prevented a defendant that developed, modified, selected, or deployed such a system from arguing that a clinician’s failure to override the system was a superseding cause that cut off liability. In short, AB 2575 would have shifted both employment and tort rules toward heightened accountability for AI-assisted clinical tools, while expressly preserving other defenses, comparative fault arguments, and evidence of reasonable precautions.

Definitions

The bill contained detailed definitions that would have governed both the Civil Code and Labor Code provisions. “Artificial intelligence” was defined as an engineered or machine-based system with varying autonomy that can infer from input how to generate outputs influencing physical or virtual environments. “Automated decision system” meant a computational process derived from machine learning, statistical modeling, data analytics, or artificial intelligence that issues simplified outputs such as a score, classification, or recommendation used to assist or replace human discretionary decisionmaking and that materially impacts natural persons. The bill excluded ordinary tools such as spam filters, firewalls, antivirus software, identity and access management tools, calculators, databases, and datasets.

The bill also defined “clinical decision support system” as an artificial intelligence system that produces a prediction, classification, recommendation, evaluation, or analysis that aids clinical decisionmaking related to timing of care, diagnosis, or treatment. It expressly excluded booking, canceling, and rescheduling appointments, reminders, patient education, previsit materials and preparation, and payment processing, so long as the system’s independent performance of those activities did not require a professional license. “Generative artificial intelligence” incorporated the meaning in Section 1339.75 of the Health and Safety Code. These definitions matter because the bill’s duties were limited to clinical decision support systems and direct patient care, rather than every kind of software used in health care operations.

Governance and Institutional Framework

The governance structure in the bill centered on the California Labor Commissioner and, secondarily, on the courts and licensing authorities already operating under existing law. On the labor side, the bill would have made the Labor Commissioner the enforcing authority for the worker-protection article. The Commissioner would have been required to investigate alleged violations, determine whether an employer retaliated or discriminated, order temporary relief to mitigate a violation or preserve the status quo, issue citations, and file civil actions. The bill specifically tied citation contest and enforcement procedures to the existing statutory frameworks in Labor Code Sections 98, 98.3, 98.7, 98.74, and 1197.1. This is a classic administrative-enforcement design: the bill would have created a rights claim, then embedded it in established wage-and-hour enforcement machinery.

On the health-care liability side, the bill did not create a new agency or licensing body. Instead, it altered the Civil Code rules that apply in civil litigation involving a clinical decision support system. It would have restricted one particular defense in tort actions but preserved other affirmative defenses, causation evidence, comparative fault evidence, and evidence showing reasonable precautions, including clear and conspicuous disclosures and disclaimers to health facilities, clinics, physicians’ offices, or group practices. The bill also preserved the authority of licensing authorities and courts under existing law, and it explicitly stated that a Labor Commissioner determination would not control malpractice or professional discipline disputes. That separation between administrative labor enforcement, civil tort litigation, and professional licensing is a central institutional feature of the proposal.

Key Focus Areas

The first key focus area was protection of human clinical judgment. AB 2575 would have declared California public policy that a worker providing direct patient care be free to use professional judgment to make assessments and decisions within the worker’s scope of practice. The bill made clear that a policy, procedure, or protocol requiring a worker to accept, defer to, or refrain from overriding a clinical decision support system would not count as one of the worker duties the statute preserved. At the same time, the bill expressly preserved duties such as meeting the applicable standard of care, acting within scope of practice, complying with specified Business and Professions Code and Title 22 obligations, exercising independent professional judgment, and following employer or agency written policies that are otherwise lawful.

The second key focus area was liability allocation for clinical AI. In a civil action against a developer, modifier, selector, or deployer of a clinical decision support system alleged to have caused harm, the defendant would not have been able to argue that the clinician’s failure to override the AI output was a superseding cause. However, the bill would still have allowed evidence on causation, foreseeability, comparative fault, and the defendant’s reasonable precautions, including disclosures and disclaimers about intended use, scope, known risks, and limitations. This design indicates that the Legislature aimed to prevent defendants from using clinician reliance as a complete liability shield while still preserving ordinary fault-allocation principles. The result would have been a more plaintiff-friendly litigation posture in cases involving AI-assisted clinical decisions.

Implementation Framework

The implementation framework was split between employment enforcement and tort litigation. For worker-rights claims, a direct patient care worker subjected to retaliation or discrimination in violation of the article would have had the right to file a complaint with the Labor Commissioner. The Commissioner would then investigate, determine whether the adverse action violated the statute, and could order temporary relief under the existing Labor Code procedures to protect the worker while the matter was pending. The bill also authorized citations and civil actions, giving the Commissioner a broad toolkit rather than relying solely on private lawsuits. Because the bill expressly said the Commissioner need not adjudicate whether the worker actually met or failed to meet the separate professional duties listed in subdivision (b)(2), the enforcement model would have focused on employer motive and retaliation, not on replacing professional-negligence tribunals.

For clinical decision support system liability, the implementation framework would have operated in court through amended Civil Code Section 1714.48. Litigants alleging harm from an AI-enabled clinical system would have litigated under the new no-superseding-cause rule, while defendants could still argue causation, foreseeability, comparative fault, and reasonable precautions. The bill did not create reporting forms, certification programs, or pre-use approvals, and it did not establish a dedicated AI regulator. Instead, it inserted AI-specific rules into existing legal systems: labor enforcement for retaliation complaints and civil litigation for harm claims. That makes AB 2575 less of a standalone AI governance regime and more of a targeted legal adjustment affecting health care decision support and workplace autonomy.

Monitoring and Evaluation

The bill’s monitoring architecture relied on ordinary enforcement rather than specialized AI audits. In the labor context, the Labor Commissioner would have monitored employer conduct by investigating complaints and assessing whether adverse action was taken because a worker overrode or relied on the output of a clinical decision support system. The statute allowed the Commissioner to consider alternate reasons for the employer’s conduct, including alleged failure to satisfy duties such as standard of care, scope of practice, or regulatory compliance. That means the agency’s fact-finding would have focused on motive, documentation, and workplace records, not on evaluating the technical performance of the AI system itself.

In the civil context, monitoring would have been indirect and case-based. The bill would have encouraged defendants to keep evidence of reasonable precautions, disclosures, and disclaimers, because such evidence was expressly preserved. As a practical matter, that creates incentives for deployers to document intended use, known limitations, and communications to clinical settings. The statute did not mandate periodic audits, incident reporting, or public transparency disclosures, so there is no built-in statewide performance dashboard. Instead, evaluation would have occurred through employer investigations, court filings, expert testimony, and ordinary litigation discovery. Since the bill was vetoed, even this indirect monitoring structure never took effect.

Penalties, Liability, and Appeals

AB 2575 would have used existing enforcement remedies rather than creating new penalty schedules. On the labor side, the Labor Commissioner could have issued citations and filed civil actions, and citation enforcement would have followed the usual procedures in the Labor Code sections referenced by the bill. The text does not set a unique penalty amount in the enrolled language, so the enforcement consequences would have depended on the incorporated Labor Code processes. The bill also contemplated temporary relief to preserve the status quo, which is significant because it allows fast administrative intervention before a full hearing is concluded.

On the liability side, the most important change would have been the removal of one defense: a defendant could not claim that a clinician’s failure to override the AI output was a superseding cause severing liability. The bill otherwise preserved defenses and fault allocation. It also included a carve-out stating that the new Civil Code rule would not apply to an action under Code of Civil Procedure Section 340.5 against a person providing direct patient care who is licensed or certified under specified professional licensing laws. The text does not provide a special appeal process for the civil liability rule. Administrative disputes would have proceeded under the ordinary contest procedures incorporated by reference. Because the Governor vetoed the bill, these liability and appeal consequences never became operative law.

Relationship to Other Instruments

AB 2575 was closely related to existing California labor, civil liability, and professional-standards law. The bill expressly preserved duties grounded in the Business and Professions Code, Title 22 regulations, employer policies, and the independent professional judgment obligation. It also preserved the authority of licensing authorities and courts under existing law, signaling that the measure was meant to coexist with medical-malpractice and professional-discipline systems rather than displace them. The Civil Code amendment built on existing principles of negligence and causation, but narrowed the availability of a superseding-cause defense in a specific AI-health-care context. Likewise, the Labor Code article relied on the Labor Commissioner’s existing enforcement architecture instead of establishing a new administrative agency.

The bill also interacted with Health and Safety Code terminology by incorporating the definition of “generative artificial intelligence” from Section 1339.75. However, the bill did not regulate all generative AI uses in health care; it targeted clinical decision support systems and direct patient care. That makes it narrower than broad AI governance laws and more specific than general technology policy. The measure would have complemented, rather than replaced, existing patient-care standards, hospital policies, and licensing rules. Since it was vetoed, however, it remains a non-operative legislative instrument, useful mainly as evidence of California’s policy direction rather than as binding law.

National/Federal Alignment

AB 2575 would have aligned California with a broader national debate over AI in health care: how to preserve human oversight while assigning responsibility for machine-assisted decisions. At the federal level, there is no single nationwide health-care AI statute in the materials relied on here that would have preempted this bill. The measure therefore appeared to operate as a state-specific tort and labor initiative, filling gaps left by general federal oversight regimes. Its insistence that clinicians remain free to exercise professional judgment is consistent with common federal and state health-care practice norms, but the bill would have gone further by creating a specific anti-retaliation protection and a litigation rule against superseding-cause defenses.

At the same time, AB 2575 would have differed from product-safety or device-approval models because it did not establish a premarket certification program, model registry, or federal-style conformity assessment. It also did not create an AI-specific federalism conflict in the text itself, because it regulated employer conduct and civil defenses within state law. The practical effect would have been strongest in California health facilities, clinics, physician offices, and group practices, where deployers of AI tools would have needed to account for state labor protections and tort exposure. As enacted? None of that became binding, because the bill was vetoed on 2026-09-30.

Implementation Timeline

Sources and References