California is writing a law that tries to answer a question the mental health system has been avoiding since the first therapy chatbot went viral: when does an AI conversation become therapy? The pending bill, SB 903, gives an answer that sounds simple and is not: if the interaction is psychotherapy, the rules for psychotherapy apply, regardless of what the company calls the product. A chatbot that calls itself a wellness coach and conducts a therapeutic session would be regulated the same as a licensed professional doing the same thing.
The need for the line is documented in the numbers behind the bill. An estimated one in eight adolescents and young adults has used a chatbot for mental health support, wrongful death lawsuits have been filed against chatbot makers, and roughly 1.2 million people per week have shared thoughts of suicide with one major consumer AI product, as Jefferson Public Radio reported. The bill's author, Senator Steve Padilla, has framed the legislation around conduct rather than labels, and that framing is the entire legal architecture: an app's marketing cannot define its way out of the duties that attach to therapy.
What the bill would do
SB 903, the Wellness and Oversight for Psychological Resources Act, has three main parts. The first is the clearest: no entity could provide, advertise, or offer psychotherapy services in California unless conducted by a licensed professional. That effectively bars companies from selling chatbot products as therapy, a market that exists today with almost no regulatory friction. The second part reaches into licensed practice: AI could not make independent therapeutic decisions, could not directly interact with clients in therapeutic communications except through an FDA-approved, HIPAA-compliant product, could not generate treatment recommendations or plans without professional review, and could not be used to detect emotions or mental states. The third part governs the administrative middle: AI may still handle scheduling, billing, records, and referrals, and patients must receive written notice and give separate written consent when AI records or transcribes their sessions, a structure Reed Smith's analysis walks through provision by provision.
Enforcement would sit with the Department of Consumer Affairs, with civil penalties up to $10,000 per violation. The bill exempts religious counseling, peer support from people with lived experience, and self-help materials that do not hold themselves out as therapy.
The bill was still awaiting a vote in the Assembly's fiscal committee as of August, which means its final shape is not settled. What is settled is the direction, and California has been moving in that direction for two years. Two earlier laws that took effect January 1, 2026 already bar AI systems from implying they hold a healthcare license and require companion chatbots to notify users they are interacting with AI, to prevent responses that could encourage self-harm, and to refer users in crisis to crisis services.
The two arguments, stated fairly
Supporters, led by the state's professional associations for psychologists and therapists, argue that the safeguards exist because the failure modes are documented rather than hypothetical. A chatbot can provide plausible but wrong clinical guidance, can mishandle a crisis, and cannot reliably detect when a user is at risk of self-harm. The National Union of Healthcare Workers, which has filed its own complaint against Kaiser Permanente over alleged automated mental health triaging, argues that AI in this setting needs the same guardrails as any clinical tool.
Opponents, led by the tech industry group TechNet, argue the bill would restrict beneficial uses of AI at the worst possible moment. Every county in California faces a behavioral health workforce shortage, and AI screening and intake tools are among the few ways the system has found to stretch scarce clinicians further. The opponents' core claim is that banning the tools does not create the clinicians, and that the clinician bottleneck the bill would protect against is smaller than the access crisis the bill would entrench.
Both arguments are stronger than the summaries they get. The workforce shortage is real and severe, and the harms the bill documents are real and severe. The bill's design does not actually choose between them; it draws a line between what AI may do without a clinician and what it may not, and both sides disagree mainly about where the line falls.
The conduct test, and why it is hard
The bill's central idea, that conduct rather than labels determines regulation, is where the legal fights will concentrate. A product that says it is not therapy while doing the work of therapy is the exact case the bill is written to catch, and the exact case the industry will argue is ambiguous. What makes an interaction therapeutic? A user who describes a bad week to a general-purpose chatbot has not necessarily received psychotherapy; a user who describes the same week to a product that responds with cognitive restructuring exercises probably has. The bill punts the definition to the enforcement process, which is honest about the difficulty and leaves the boundary to be drawn case by case.
The disclosure rules are where the bill is least controversial and most likely to survive unchanged. Requiring notice that the user is interacting with AI, and consent before sessions are recorded or transcribed, tracks what patients already expect from telehealth and what the January laws already require of companion chatbots.
What it means for the people using these tools
For the roughly one in eight young people who have talked to a chatbot about their mental health, the bill changes what the product may claim rather than whether it exists. A chatbot may still talk. It may not claim to treat. And if its makers market it as a therapist, California intends to hold them to the standard of one, which is precisely the standard none of the current products could meet.
The harder truth sits underneath the legislation. People turn to AI for mental health support in large part because human care is expensive, scarce, and hard to reach, and a law that restricts the AI alternative does not by itself produce a single additional appointment. The bill's defenders acknowledge this and answer that the response to a shortage is not a product that cannot recognize a crisis. That exchange, more than any committee vote, is the real subject of SB 903, and whatever California decides will become the template other states either copy or reject.
If you or someone you know is in crisis, the 988 Suicide and Crisis Lifeline is available by call or text at 988.
The laws already in force set the pattern
SB 903 is the third act of a policy California began two years ago, and the earlier acts show how the state intends to enforce the line it is drawing. AB 489, in force since January 1, 2026, prohibits AI systems from using any terms, letters, or design elements that imply the AI holds a healthcare license, and bars advertising that suggests care is provided by a licensed person when it is not. The licensing boards themselves gained jurisdiction to pursue injunctions, which is a notable enforcement choice: the boards that regulate therapists are the ones empowered to go after the chatbots impersonating them.
SB 243, also in force since January, regulates companion chatbots designed for ongoing interaction and emotional support. It requires clear notice that the user is talking to AI, mandates protocols to prevent responses that could encourage self-harm, and requires referral to crisis services when a user expresses suicidal ideation. The statute treats the crisis moment as the regulatory priority, which tells you what the legislature worries about most: not that a chatbot is mediocre, but that a chatbot is the only thing listening when someone is in danger.
The Kaiser Permanente complaint filed by the National Union of Healthcare Workers shows the same logic reaching into established institutions. The union alleges Kaiser used automated mental health triaging in ways that disadvantaged patients, and the complaint has become part of the bill's supporting record. The pattern across all three measures is consistent: AI may assist care, AI may not substitute for it, and the moment of crisis is where the substitution is most dangerous and most regulated.
That pattern is also why the workforce shortage argument carries only so far. The bill's opponents are right that California lacks clinicians, and the bill's supporters are right that a crisis-incompetent chatbot does not fill the gap. The law's actual answer is to permit AI for the administrative and supplementary work that stretches clinicians, scheduling, records, referrals, and to reserve the therapeutic interaction itself for humans. Whether that division is workable in a system this short of humans is the empirical question the next decade of California healthcare will answer.
The bill's final form will be settled in the Assembly, and the version that emerges will matter far beyond California, because the state's healthcare market is large enough that its rules become de facto national standards for app makers. A company that wants to sell an AI mental health product in the United States will not build two versions of its crisis handling; it will build one that satisfies California. That is the quiet mechanism by which this one bill regulates the whole country, and it is the reason the conduct test is worth watching: whatever line California draws between an assistant and a therapist is the line every other state will inherit.
Primary sources
- Jefferson Public Radio reporting on the bill, the usage statistics, and the lawsuits behind it.
- Reed Smith analysis of SB 903's provisions and exemptions.
- State Affairs Pro coverage of the bill's status and the political dynamics.
- Lexology summary of the January 2026 laws on AI healthcare licensing and companion chatbots.