How SB 903 Changes AI Use in California Therapy Practices

As of September 2026, SB 903 has passed the California Senate unanimously, cleared two Assembly committees, and awaits Governor Newsom's signature. The bill targets a mismatch: consumer AI chatbots are marketed as therapy while licensed care is delivered by humans.

That mismatch defines the tension for counselors. AI can widen access and reduce administrative burden, but practitioners who adopt it carry risk when a consumer tool is mistaken for care. SB 903 responds by banning advertising claims that a chatbot is therapy and forbidding AI from making unsupervised clinical decisions, with penalties tied to those categories. The wrongful death lawsuits moving through California federal courts make that risk concrete.

What SB 903 Actually Does

Does California now ban AI therapy chatbots, or is the fight still playing out? As of early September 2026, SB 903 is pending before Governor Gavin Newsom, but it has already cleared the Legislature with broad support.¹

Legislative status

Authored by Senator Steve Padilla of San Diego, the bill would prohibit companies from advertising AI therapist chatbots as therapy and set standards for licensed providers' use of AI in mental health care. It passed the Senate 39-0 on May 19, 2026, cleared Assembly policy committees, and was later released from the Assembly Appropriations suspense file with a 13-0 vote on August 13.¹ The Assembly passed the amended version on August 30, and the Senate concurred in those amendments 40-0 on August 31.¹ Therapist Support's tracking page notes the earlier progress and the wrongful death litigation surrounding AI chatbots.¹ The Governor has until September 30, 2026, to sign, veto, or let it become law unsigned.¹ No action has been announced as of this writing.

Telehealth and out-of-state uncertainty

SB 903's public framing targets businesses "offering or advertising therapy or psychotherapy in California," but available text does not spell out how that applies to out-of-state or telehealth providers serving California residents, leaving cross-state telehealth counseling laws unclear. Counselors licensed in another state but treating California clients should treat the bill as potentially relevant and monitor board guidance rather than assume an exemption.

The bill does not ban all AI use by clinicians. It targets advertising claims and clinical decision-making, leaving room for tools that support, rather than replace, licensed judgment.

What Counts as a 'Companion Chatbot' Under SB 903

SB 903 uses a statutory definition that zeroes in on the design of the tool, not the word "chatbot" in a product name. To count as a "companion chatbot" under the bill, an AI system must have a natural-language interface and generate adaptive, human-like responses. It must also be capable of meeting social or emotional needs, which can include anthropomorphic features such as a name, avatar, or persona. Finally, the system must be able to sustain a relationship across multiple interactions, so a single turn or scripted menu does not qualify.

Qualifying and non-qualifying examples

• Likely in scope: A friendship-style chatbot that remembers prior conversations, checks in with the user, and maintains an ongoing persona. An emotional-support bot that carries on human-like back-and-forth conversations would also qualify, even if it never describes itself as a therapist or medical product, a dynamic also relevant to digital self-help therapy.
• Likely outside scope: A customer-service chatbot that helps with a billing or scheduling question is not a companion chatbot under the statute.¹ The same is true for internal tools used for documentation, research, or technical help, for video-game chatbots limited to game-related interactions, and for stand-alone voice assistants that do not sustain an emotional relationship.

Why the definition matters for advertising

The companion-chatbot definition is the legal trigger for the advertising restrictions. If a product meets this definition, its maker cannot advertise it as therapy or otherwise imply that it provides clinical mental health care. By contrast, a scheduling assistant or clinical notes tool falls outside the definition, so it is not swept into the therapy-advertising ban. For counselors and social workers, the practical rule is to look at how the system functions in real sessions and between sessions, especially whether it forms an ongoing emotionally engaged relationship, rather than relying on the vendor's label, an approach consistent with broader changes in technology in counseling.

What's Banned: Advertising and Clinical Decision-Making

California's SB 903 creates two categories of prohibited AI activity in mental health: advertising a chatbot as therapy and allowing AI to make unsupervised clinical decisions.¹

Advertising Restrictions

Under the bill, companies and practitioners cannot advertise an AI chatbot as providing psychotherapy or other mental health therapy. Claims that imply a licensed professional is delivering care when the AI is doing the work are prohibited, including use of protected professional titles in AI product marketing. A companion chatbot cannot be marketed as a substitute for therapy. By contrast, a permitted disclaimer clearly labels the product as a non-clinical support tool and states that it does not diagnose, treat, or replace a licensed professional. The key distinction is whether the message claims or implies a therapeutic relationship and clinical judgment by the AI.¹

Clinical Decision-Making Restrictions

SB 903 bars AI from making independent therapeutic decisions. AI cannot generate a diagnosis, assess symptoms for triage, or produce a treatment plan without review by a licensed clinician. It also cannot detect emotions or mental states, and it cannot directly interact with clients therapeutically unless the product is FDA-approved for low-risk general wellness or clinical support software and meets HIPAA standards. Session recording or transcription is permitted only after the patient is informed and consents.¹

A noncompliant workflow looks like this: a user tells an app they feel hopeless, and the app responds with a depression label and a medication recommendation without any clinician review. That violates both the symptom assessment ban and the independent clinical decision ban. A compliant workflow would flag the statement, then route it to a licensed professional who reviews and decides next steps. Providing AI therapy to the public remains the responsibility of a licensed professional.¹

What AI Uses Remain Permitted for Licensed Clinicians

SB 903 is not an outright ban on AI in mental health care, and licensed counselors should not read it that way. The bill targets unregulated advertising and automation of clinical judgment, but it leaves meaningful room for AI tools that operate under a licensed clinician's supervision.

Permitted Clinical and Administrative Uses

Licensed providers may still use AI for several routine functions. Session transcription and dictation can turn audio into searchable progress notes. Note-taking assistants can organize themes, symptoms, and treatment goals before the clinician reviews them. Administrative triage tools can route appointment requests, flag missed sessions, or summarize intake forms. Clinician-supervised decision support can surface questions, risk flags, or evidence-based resources, as long as the software does not independently make the call.

A Compliant Workflow Example

A common compliant workflow looks like this: after a session, the therapist uses an AI tool to draft a progress note from the session transcript. The clinician then reviews the draft, corrects any errors, adds clinical nuance, and signs the note in the EHR. The AI may also highlight follow-up tasks, such as a pending release of information or a symptom pattern to reassess, but the therapist decides whether any suggestion is accurate and appropriate.

The Decision-Maker of Record

The core requirement is that the clinician remains the decision-maker of record. SB 903 does not prohibit AI support; it prohibits shifting therapeutic authority to a chatbot. In documentation, clinical supervision, and consultation, counselors should be able to show they evaluated the AI output rather than adopting it automatically. That distinction, between a tool that assists and a product that acts as therapist, is the line California is drawing.

Related Articles

Penalties and Enforcement Under SB 903

SB 903 establishes civil penalties for unlawful AI mental health therapy products and advertisements, while licensed professionals may face separate board discipline. The table below shows the violation types, penalties, and responsible authorities identified in the bill and its legislative analyses.

Consent, Disclosure, and Recordkeeping Requirements

In 2026, California's SB 903 would make verbal or written disclosure and explicit consent mandatory before any AI tool records, transcribes, triages, or screens a psychotherapy client. That requirement sits alongside, not in place of, existing California confidentiality law.

What providers must disclose before AI use

Before activating an AI note-taker, transcription service, or triage assistant, clinicians must tell the client or the client's legally authorized representative that AI will be used and explain the specific purpose. Consent must be explicit, informed, and voluntary. It must be documented in the client record, and the client may revoke it. Generic website terms or passive behavior, such as not objecting, are not enough. The same consent framework applies to AI-assisted triage and screening, not just recording or transcription. The disclosure should occur before the AI tool encounters any protected information.

How the AI consent layer fits with CMIA

SB 903 adds requirements on top of the Confidentiality of Medical Information Act (CMIA); it does not replace or weaken CMIA. Once an AI-generated transcript or recording exists, it is protected medical information under CMIA. Any vendor or third party that handles those records must comply with CMIA's release and confidentiality restrictions, even after a client has consented to AI use. For psychotherapy records, CMIA requires written disclosures about intended uses and retention periods before information is released to third parties.

Recordkeeping obligations

At a minimum, the consent itself must be documented in the clinical record. Documentation should show the date, the person providing consent, and the specific AI functions authorized. SB 903 does not create a separate AI-specific retention schedule, so providers should continue to follow CMIA's existing retention and disclosure rules. If AI-generated transcripts or recordings are shared with a vendor or other third party, keep a record of the CMIA-required written disclosures, including intended uses and retention period.

Client-facing vs. clinician-facing tools

The consent duty attaches to the AI function, not the interface. A clinician-facing ambient scribe that records a session triggers the same disclosure and consent requirement as a client-facing chatbot that triages symptoms. Refusing AI consent does not affect a client's right to receive care.

The standard of care is shifting: when AI is used, the licensed clinician, not the vendor, remains responsible for disclosure, consent, and the clinical decision itself.

California SB 903 compliance guidance

Ethical and Liability Implications for Practice

SB 903 is a direct legislative response to wrongful death litigation already unfolding in California federal court, and that litigation is now shaping the liability landscape for clinicians who use AI at any point in care.

Wrongful Death Cases Enter the Courtroom

The most concrete case referenced in SB 903 coverage is Lyons v. Open AI Foundation, case number 3:25-cv-11037, pending in the U.S. District Court for the Northern District of California before Judge William H. Orrick and filed in 2025 as a wrongful death action.¹ Separate reporting describes a second California wrongful-death suit against OpenAI alleging that GPT-4o contributed to a murder-suicide, though that case has no publicly identified caption or filing date. Together these filings signal that AI-related harm is being treated as a real, justiciable injury, not a hypothetical risk.

A New Standard of Care for AI-Assisted Decisions

For counselors, therapists, and social workers, the practical effect is that courts and licensing boards may start asking whether a reasonable clinician following the APA Code of Ethics or the BACB Ethics Code in the same situation would have relied on a particular AI tool. If a clinician uses an unvetted consumer chatbot, fails to verify its output, or lets the tool make a clinical call, the standard-of-care question becomes harder to defend. SB 903 does not have to create a private right of action for these cases to matter. A statutory violation can still be introduced as evidence that a provider fell below accepted professional norms.

Documentation That Protects Clinicians

Because AI-assisted decisions may be scrutinized after an adverse event, records should show the AI was an adjunct, not the decision maker. Document the tool and version used, the output or summary reviewed, the independent clinical reasoning that accepted or rejected that output, and the client's consent to any AI involvement. When the record makes the clinician's judgment the final step, it is far easier to show that the standard of care was met.

How Counselors Can Prepare: A Compliance Checklist

Use this checklist to align your practice with the requirements taking shape under SB 903, while recognizing that final implementing regulations may still evolve.

1. Audit current AI tools

   Review any chatbots, AI companions, or automated intake tools you currently use or recommend against the bill's companion chatbot definition. When in doubt, treat unapproved consumer apps as out-of-scope for clinical use.
2. Update informed consent forms

   Clearly disclose if AI is used for recording, transcription, triage, or other parts of care. Written consent should explain what the tool does, what it does not do, and when a human clinician reviews output.
3. Train staff on prohibited advertising language

   Teach front-desk, marketing, and intake staff to avoid describing AI tools as therapy, counseling, therapeutic, or as a provider of mental health treatment.
4. Document clinician review and sign-off

   For any AI-assisted clinical note, recommendation, or triage suggestion, include a dated record showing a licensed clinician reviewed and approved it before it entered the client record.
5. Verify vendor compliance claims

   Request written documentation from AI vendors about regulatory compliance, data handling, and clinical decision support. Do not rely on marketing language alone.
6. Set a recheck reminder

   Place a calendar reminder to review SB 903's final status and any implementing regulations, since enforcement details and definitions may shift before full implementation.

Explore California Resources