California's new rule for AI mental-health chatbots: support, don't practice

A user types a message about feeling anxious before a job interview. The chatbot, called "Psychologist," responds with italicized body language cues, including "nods slowly," and reassurance in the voice of a seasoned therapist. Only a small-font disclaimer at the bottom of the screen reads: "This is A.I. and not a real person. Treat everything it says as fiction."

That disclaimer, and the marketing category it is trying to soften, is exactly what California state Sen. Steve Padilla wants to outlaw. Senate Bill 903, the bill he introduced this month, does not try to ban AI mental-health chatbots. It tries to redraw a specific legal line: between AI tools that "support" clinicians, and AI systems that "practice" psychotherapy.

"AI can support clinicians," Padilla said in announcing the bill. "It cannot practice psychotherapy."

The bill does three concrete things. It would bar companies from advertising their chatbots as therapy, or marketing them in ways that imply a licensed clinician is on the other end. It would prohibit any AI from making therapeutic decisions, including diagnoses, treatment choices, and risk assessments, without review by a licensed human professional. And it would require health providers to disclose, and get patient permission for, any use of AI to record therapy sessions or to triage mental-health care, according to the bill text summarized in the Los Angeles Times.

The provisions are narrow on their face. They are not aimed at a chatbot someone chats with for fun late at night, or at a productivity assistant bolted onto a clinic's scheduling system. They are aimed at a market that has started to look, and advertise, a lot like clinical care without any of the licensing, supervision, or liability that clinical care carries.

Padilla's coalition gives the bill weight. It is backed by professional associations representing psychologists, therapists, and counselors, and by the National Union of Healthcare Workers, which represents mental-health clinicians across the state. NUHW has spent the summer sharpening that argument. In July, the union filed a complaint with state regulators against Kaiser Permanente alleging the health system used an automated algorithm to triage mental-health patients, a practice the union says short-circuits clinical judgment.

The bill's moral urgency comes from a separate, faster-moving track: wrongful-death lawsuits filed in California federal courts against the makers of AI companion chatbots, alleging the products contributed to user suicides. Padilla has pointed to those suits as the reason the rule is needed now, while the products are still scaling.

Roughly one in eight teens and young adults are estimated to be turning to chatbots for mental-health help or advice, a usage figure large enough that even a small failure rate lands a lot of people in the gap the bill is trying to close.

The line Padilla is drawing, though, runs straight through the word "support." A clinician who uses an AI tool to draft session notes is supporting care. A chatbot that walks a patient through a cognitive-behavioral exercise is doing something closer to practice, and depends on who is supervising it. SB 903 does not yet define where "support" ends and "practice" begins, and that ambiguity is the place the next round of lobbying will land.

The practical question for patients, in the meantime, is simpler: ask whoever is on the other side of the screen whether a licensed human is reading, signing off, or sharing responsibility for what the AI is telling you. Under SB 903, providers would have to answer.