California’s SB 903 Targets AI Therapy Claims and Clinical Use

The distinction matters because the evidence around chatbot support is still unsettled. A report on the legislation cited RAND researcher Jonathan Cantor’s warning that there are few standardized benchmarks for mental-health advice from AI chatbots and limited transparency about the data used to train large language models. That is a limitation on claims of clinical equivalence, not proof that every nonclinical use is harmful.

The bill’s operational burden sits with providers and deployers

SB 903 would permit a person or organization that provides or facilitates psychotherapy to use AI for administrative or supplementary support. Its examples include appointment management, billing and insurance claims, record maintenance, workflow support, professional-reviewed progress analysis, and organizing resources or referrals.

The bill treats recording, transcription, triage and screening differently. Before those uses, the patient or an authorized representative would have to be told that AI will be used and the tool’s specific purpose, then give consent. The statutory definition requires a clear, affirmative, informed and voluntary act that is documented and revocable; a broad terms-of-use agreement would not qualify. It also says a patient does not surrender a right to care by declining consent for AI recording, transcription, triage or screening.

For AI used in connection with psychotherapy, triage or screening, the measure would require licensed review and approval for therapeutic decisions, therapeutic recommendations, assessments, diagnoses, treatment plans, emotion or mental-state detection, and triage or screening. It separately limits direct psychotherapeutic communication, unless a tool is cleared or approved by the Food and Drug Administration for that use and complies with HIPAA.

Responsibility would follow control. A licensed professional using a tool would have to ensure clinically appropriate, compliant use by that professional and those under their supervision. If an employer or contractor requires or authorizes the tool, the employer or contractor would be responsible for compliant deployment and direction. The bill would also require psychotherapy records to comply with California medical-information confidentiality rules and would prohibit sharing, selling, storing or training models on psychotherapy data in a way inconsistent with applicable law.

That allocation is the bill’s larger policy move. It does not make a developer the sole focus; it places compliance duties on the clinician, employer or contractor that chooses and deploys the system. Relevant licensing boards or enforcement agencies could pursue existing remedies, including injunctions or restraining orders, and the boards could issue implementation rules.

The access argument cuts both ways

The proposal’s supporters include the California Psychological Association, the California Association of Marriage and Family Therapists, the California Behavioral Health Association and the National Union of Healthcare Workers, according to Padilla’s Senate-passage announcement. Their shared stake is preserving professional accountability where tools could influence vulnerable patients’ care.

Le Ondra Clark Harvey, chief executive of the California Behavioral Health Association, told lawmakers that inaccurate chatbot information or a mishandled crisis can be life-altering. The National Union of Healthcare Workers has pursued the same concern through a complaint against Kaiser Permanente’s e-visit screening tool.

That complaint illustrates the bill’s difficult application rather than resolving it. The union says the tool asks people concerned about anxiety or depression a multiple-choice questionnaire and automatically generates care recommendations and referral pathways; it argues that the speed of those recommendations makes licensed review unlikely. Kaiser told the report that the tool does not use AI to diagnose patients, make clinical decisions or determine medical necessity; state regulators are investigating the complaint. The same report says it is unclear whether SB 903 would apply to that system.

TechNet’s stated objection centers on the clinical-use provisions. Robert Boykin, its California executive director, argued that the proposal puts a clinician bottleneck ahead of intake and screening tools intended to get patients to care faster. The comparison is not between an unregulated consumer chatbot and a clinician; it is between an automated workflow with mandated professional review and one that can operate with less of it.

California Legislative Information’s status page lists SB 903’s five most recent legislative actions. Source: California Legislative Information bill history.

The next test is whether the terms can be made workable

The Senate passed SB 903 39–0 on May 19. The official legislative history records Assembly amendments on July 2 and, on August 5, placement on the Appropriations Committee’s suspense file.

Before the bill can advance, lawmakers will have to decide whether its definitions make the dividing line operational. The text defines triage or screening as assessing health concerns and symptoms to determine urgency, clinical nature or the appropriate level of psychotherapy need. It leaves implementation to future regulation as well as the bill’s wording. The unanswered evidence is whether providers can apply the consent, review and responsibility rules consistently to real intake systems—and whether those rules catch unsafe automation without treating ordinary administrative support as psychotherapy.