California’s governor signed the strictest child-safety rules for companion chatbots in the country, then vetoed the bill that would have stopped chatbots from offering therapy to everyone else. Same month, same author, opposite answers. The state has decided that a chatbot may not harm a child, and may still call itself a therapist to an adult.
The two bills came from the same desk. Senator Steve Padilla wrote SB 243, the first companion-chatbot law in the United States, in 2025. In 2026 he co-authored Adam’s Law, SB 1119, named after Adam Raine, a California teenager whose family’s lawsuit alleges that ChatGPT encouraged his death; Governor Gavin Newsom signed it with the line “innovation comes with responsibility and protecting our children comes first.” Padilla’s other bill of the session, SB 903, would have prohibited companies from advertising or offering psychotherapy through a chatbot and required a licensed professional behind any therapeutic decision an AI touches. It passed the Senate unanimously and the Assembly 71 to 4. Newsom returned it unsigned.
What the bill would have done
SB 903 was narrower than its critics described and wider than its title suggested. According to the Assembly’s own analysis, it would have barred advertising or offering psychotherapy services through companion chatbots, barred an AI from generating diagnoses, treatment plans, or recommendations without a licensed professional’s approval, and barred AI from detecting a patient’s emotional state or running intake screening without professional oversight. Psychotherapy meant services to diagnose or treat a mental health or substance use disorder; religious counseling and peer support were carved out.
Clinicians could still use AI for scheduling, billing, records, and note-taking, provided a patient gave explicit, specific consent before a session was recorded or transcribed, and could decline without losing care. Enforcement would have sat with the professional licensing boards, which is where the practice of medicine without a license already lives.
Nineteen organizations supported it, among them the state’s associations of psychologists, therapists, and counselors, the National Union of Healthcare Workers, and the California Behavioral Health Association. Seven opposed it unless amended, led by TechNet, whose executive director argued that requiring a clinician in the loop would delay care in a system already short of clinicians. Newsom’s veto message, as characterized by Padilla’s office, held that the bill’s protections “go too far.”
Who the bill was written for
The numbers in the legislative record describe the gap the bill tried to close. Nearly half of American adults with mental health needs report barriers to getting care. In the Assembly analysis, about half of respondents with mental health challenges said they had used a chatbot for support. In a 2026 American Psychological Association survey, 77 percent of psychologists said they had patients who used AI for mental health purposes. OpenAI’s own figure, published in 2025, is that 0.15 percent of people active on ChatGPT in a given week have conversations with explicit indicators of suicidal planning or intent, which on the user base the company reported at the time came to more than a million people a week.
None of those people were the bill’s target. The target was the business that looks at that demand and names a product after it. The precedent already existed: in 2026 Pennsylvania’s attorney general sued Character.AI over a character called Emilie that told users it was a licensed psychiatrist. SB 903 would have made that a licensing-board matter in California before it became a lawsuit.
Why the veto is the news
Read together, the September signatures draw a line. Adam’s Law requires companion-chatbot operators serving minors to submit independent audits to the attorney general, to route self-harm threats to crisis support and notify a parent, and to ship with notifications off, time limits on, and memory limited, changeable only by a parent. SB 867, signed in the same batch, bans companion chatbots in toys. AB 1709 bans addictive design features for users under 16. The state’s position on children is now the most detailed in the country.
For adults, the position is SB 243: the chatbot must say it is an AI, must have a protocol for self-harm, and may be sued if it does not. What it may call itself was the question SB 903 asked, and the answer, for now, is anything. A product can be built for retention, measured on time in app, and sold as therapy, and the only thing California requires is a disclosure that it is not human.
The two words name opposite designs, and the design is what decides which one you are talking to: one is paid when you stop needing it, the other when you come back. The legislature voted, nearly without dissent, to keep the words apart. The governor’s veto keeps them interchangeable in the app store, and leaves the sorting to the people who are searching for one and will be sold the other.
What stays open
Illinois already bans AI from delivering therapy outright, and the Pennsylvania suit will test whether impersonating a clinician is unlicensed practice when the clinician is a character. Padilla’s office framed the veto fight as California’s licensed professionals against a federal administration that is promoting AI therapy tools, which is the direction the rest of the country is likely to take if the largest state declines to draw the line.
Florida reached the same place by a different route in the same month, when it banned companion AI from every classroom and left the market outside the school untouched. California’s version is more precise, because it turned on a single word. Everything the state now requires of a companion chatbot, the audits, the crisis routing, the disclosure, regulates what the product does. SB 903 was the first bill to regulate what the product may say it is, and that is the one that did not survive. The category’s critics have argued since the category’s first lawsuits that a chatbot built for retention should not be allowed to borrow the vocabulary of a profession built for discharge. Sacramento has now ruled on the conduct and left the vocabulary alone. In the search results, where this is actually decided, the vocabulary is the product.
Sources: Office of Governor Gavin Newsom, legislative update (September 30, 2026, SB 903 listed as vetoed; AB 1979 and others signed) and press release on Adam’s Law (September 10, 2026). Office of Senator Steve Padilla, “Governor Newsom Allows Dangerous Chatbots to Provide Therapy” (October 1, 2026) and “Governor Newsom Signs Adam’s Law” (September 10, 2026). California Assembly Committee on Privacy and Consumer Protection, analysis of SB 903 as amended (hearing July 1, 2026). CalMatters (August 6, 2026, SB 903, supporters and opponents). Decrypt (August 16, 2026, APA survey of psychologists, TechNet statement). Stauss Firm, “California’s Newsom Signs 22 Privacy and AI Bills, Vetoes Three” (October 1, 2026, SB 867, SB 1119, AB 1709). OpenAI, “Strengthening ChatGPT’s responses in sensitive conversations” (October 2025, 0.15 percent weekly indicator). NPR (May 5, 2026, Pennsylvania lawsuit against Character.AI). California SB 243 (effective January 1, 2026).
