Technology

California bill draws line on AI therapy

By Staff Writer  |  11 August 2026

A California bill would stop companion chatbots being advertised as therapists and require licensed review before artificial intelligence makes clinical recommendations or decisions.

Senate Bill 903, amended in the state Assembly on 2 July, would confine the use of artificial intelligence in psychotherapy to administrative or supplementary support. It would also regulate recording, transcription and initial assessment when a machine is involved.

Patients would have to be told that artificial intelligence will be used and why. Consent would need to be explicit, voluntary, recorded and capable of being withdrawn. Refusing consent could not be treated as surrendering a right to care.

The bill does not remove artificial intelligence from clinical work; it places a licensed professional between the system and the decision.

Chatbots could not call themselves therapists

The proposed law would prohibit companies from advertising psychotherapy delivered through companion chatbots. It would also stop a system from making therapeutic decisions, detecting emotional states or producing diagnoses and treatment plans without professional review and approval.

AI algorithms are not fit to take over the job of human therapists, who have skills and training that AI is incapable of replicating.

Steve Padilla, California State Senator

The bill allows narrower uses, including scheduling, billing, record preparation and progress analysis reviewed by a professional. It also permits tools that organise external resources or support documentation, provided responsibility for clinical decisions remains with the licensed person.

Data use is another dividing line. Records involving artificial intelligence would remain subject to medical confidentiality rules. A company could not share, sell, store or train models on psychotherapy data in a way that conflicts with applicable law.

A test for health technology rules

The measure arrives as more people use general chatbots for emotional support, often outside any regulated health service. The accessibility is part of their appeal, especially where human treatment is expensive or hard to obtain.

That demand also exposes the central problem. A fluent response can appear clinical even when the system has no professional duty, no complete patient history and no reliable way to recognise a dangerous change in condition.

California's proposed answer is direct: software may assist care, but it cannot quietly become the clinician.

The drafting question is where assistance ends and practice begins. A system that summarises notes for a clinician sits plainly on one side of that line, and one that responds directly to a patient in distress sits on the other. Most products in the market sit somewhere between the two.