California Draws New Lines Around AI in Patient Care, Even as Washington Pushes Back

For the nurse at the bedside or the patient scrolling through a health app at midnight, the question of who, or what, is making decisions about their care has stopped being theoretical. This week, California answered part of that question. Gov. Gavin Newsom signed three bills into law that draw firmer boundaries around how artificial intelligence can be used in clinical settings, while vetoing two others that would have gone further still.

Think of these laws as guardrails on a highway that's being built while cars are already driving on it. AI tools are already embedded in hospital workflows, chatbots, and decision support systems across California. The new laws don't stop that traffic. They try to make sure the road has lane markers before someone gets hurt.

The centerpiece is AB 1979, sponsored by the California Nurses Association, which amends state healthcare law to keep AI in an advisory role rather than a decision-making one. It extends California's medical confidentiality protections to cover health information patients share with chatbots, and it blocks AI from performing licensed clinical functions, though clinical documentation and routine administrative communication are exempted. CNA President Sandy Reding, RN, called the law a win built on collective action. "Nurses won this fight the way we win every fight: Together," she said in a statement. "We stood up and said that AI must support, not replace, the clinical judgement of the nurses and other healthcare professionals who care for our patients."

Assembly Member Mia Bonta, D-18th District, who introduced the bill, framed it in simple terms. "Technology should not replace human decision-making in healthcare," she said, adding that the law ensures medical records "stay protected when accessed through direct-to-consumer health chatbots" and that clinicians retain authority over how AI-informed decisions get used.

A second law, SB 503, tackles a quieter but equally consequential problem: algorithmic bias. If a clinical decision support system, the kind of software that flags risk scores or suggests treatment pathways, carries a foreseeable risk of biased outcomes, developers now have to make reasonable efforts to identify that risk and mitigate it. Providers and other deployers are required to monitor these systems on an ongoing basis, not just at launch. The idea is straightforward: an algorithm trained on skewed data can quietly reproduce the very disparities healthcare systems are supposed to be correcting, and nobody notices until patients are already affected.

The third new law, AB 1609, deals with something more visible to ordinary consumers: chatbots pretending to be human. It bars large companies from representing a customer service bot as a person, but carves out exemptions for hospitals and healthcare facilities handling communications tied to the "provision, coordination, management, administration, payment or operation of health care services." In practice, that means a hospital's AI-driven scheduling or billing assistant doesn't have to announce itself as a machine every time, but the broader principle of transparency still applies elsewhere.

What Newsom Rejected

Two other bills didn't survive the governor's desk, and the reasoning behind those vetoes says as much about the politics of AI regulation as the signings do.

AB 2575, also backed by CNA, would have prevented AI developers and providers from using a clinician's failure to override an AI system's output as a legal defense in malpractice cases. It also would have expanded employment protections for healthcare workers navigating AI tools. Reding's response was pointed. "The governor knows AI should never replace a nurse's clinical judgment, but he vetoed the bill that would have protected us from using our judgment," she said.

The other veto, SB 903, targeted AI's role in mental healthcare specifically. It would have barred AI from performing therapeutic functions in psychotherapy, limiting it to administrative or supplementary support only. Given the rapid rise of AI chatbots marketed as mental health companions, often with little clinical oversight, the veto leaves a regulatory gap in one of the more sensitive corners of digital health.

None of this is happening in a vacuum. In December 2025, President Trump signed an executive order calling for a "minimally burdensome national AI standard," arguing that state-by-state regulation creates a patchwork that stifles innovation and could even force AI systems to produce "false results" in the name of fairness. The order specifically cited Colorado's AI law, set to take effect in January, which includes healthcare-specific provisions similar to California's.

That tension between state experimentation and federal preference for uniformity is already a headache for the companies building these tools. Leigh Burchell, vice president of policy and public affairs at Altera Digital Health and chair of the HIMSS Electronic Health Record Association, put it bluntly earlier this year: "From a purely kind of code perspective, it's challenging, because instead of having one certification program to manage, we have all these different things coming from different states." Multiply California's three new laws across a dozen other states with their own rules, and the compliance puzzle for health IT vendors grows quickly.

Why It Matters

None of these laws solve the underlying challenge, which is that AI in clinical decision-making is still a work in progress. Matt Cybulsky, managing director of healthcare at Catalant, noted recently that "trustworthiness still reigns supreme in adoption and delivery." Laws like AB 1979 and SB 503 are an attempt to build that trust into the system before adoption outpaces oversight, not after. Whether that approach survives a push toward federal preemption is still an open question, but for now, California nurses, patients, and developers are operating under a clearer, if incomplete, set of rules. The vetoes suggest the state isn't willing to go as far as some advocates wanted, at least not yet.