California's companion-chatbot law contains one requirement anyone can picture. If a known minor keeps chatting, the platform must remind them at least every three hours to take a break and say again that the chatbot is artificial, not human.
The timer will attract attention because users can see it. It is also the simplest part of SB 243 to build and audit. The difficult work sits behind the conversation: a self-harm protocol, a public account of that protocol, controls for minors, and an eventual reporting trail.
SB 243 took effect with California's 2026 statutes. Read as a product specification, it assigns extra duties to chatbots designed to sustain a social relationship.
Which chatbots count as companions?
The law does not classify every conversational interface as a companion. Its definition covers a natural-language AI system that provides adaptive, human-like responses, can meet a user's social needs, exhibits anthropomorphic features, and can sustain a relationship across multiple interactions.
Its exclusions sharpen the line. Customer-service bots, productivity tools, technical assistants, internal research systems, and systems that analyze source information are outside the definition. The same goes for limited game characters and voice assistants that neither sustain relationships nor generate responses likely to elicit emotion.
Product behavior matters more than the label in the navigation bar. An "assistant" that remembers a persona, seeks emotional attachment, and carries a social relationship across sessions may fit the definition. A fluent support bot does not become a companion merely because it replies in complete sentences.
For teams that route models through a shared layer such as ISH API, the finished product still determines the scope decision. A provider name cannot answer what the interface, memory, and engagement design do together.
The notice and the protocol solve different problems
The disclosure rules are direct. If a reasonable person could be misled into thinking a companion is human, the operator must clearly state that it is artificial and not human. A user the operator knows is a minor must receive an AI disclosure. During a continuing interaction, the minor must also receive a break reminder and another artificial-not-human notice at least every three hours.
Those notices address confusion and prolonged use. They do not tell the product how to respond when a conversation turns toward suicide or self-harm.
Before a companion can engage with users, its operator must maintain a protocol that prevents production of suicidal-ideation, suicide, or self-harm content. The protocol must include a crisis-service referral notification when a user expresses those subjects. The operator must publish details of the protocol on its website.
A functioning protocol needs detection criteria, response rules, escalation paths, referral behavior, testing, and failure review. It also needs careful evidence. Engineers must be able to reconstruct whether the protocol ran without turning intimate conversations into a permanent archive.
Memory adds another trap. A companion may store a preference or a prior emotional statement, but stored text is not automatically verified evidence about the user. BLOGish's guide to keeping agent preferences separate from factual evidence applies directly when an old statement could affect a later safety decision.
Reporting creates a public trail
Starting July 1, 2027, operators must report annually to California's Office of Suicide Prevention. Reports must include the number of crisis-service referral notices and the protocols used to detect, remove, and respond to suicidal ideation. They cannot include user identifiers or personal information. The office must publish data from the reports.
The referral count will not produce a clean safety ranking. A larger service may issue more notices because it has more users. A cautious detector may issue more than a narrow one. Product mix, age distribution, and conversation length can change the denominator. The statute requests a count, not a normalized benchmark.
Operators cannot keep every detail of their self-harm procedures private while providing no indication of whether interventions run. The published record gives regulators, researchers, and the public something concrete to question.
Implementation also carries civil consequences. A person injured by noncompliance can seek an injunction, the greater of actual damages or $1,000 per violation, and reasonable legal fees and costs.
California rejected a broader restriction
The state considered a more restrictive approach. In his veto message for AB 1064, Governor Gavin Newsom wrote that the bill might unintentionally produce a total ban on conversational AI for minors. He pointed instead to SB 243's notices, break reminders, sexual-content restrictions, and self-harm protocols.
The narrower choice leaves room for criticism. A three-hour interval does not prove that a minor is safe. A total prohibition could block benign educational or creative uses without examining how those products behave. SB 243 defines the covered relationship, requires controls for specific risks, and exposes parts of the operation to later scrutiny.
The Federal Trade Commission is collecting related information, but it has not announced findings. Its companion-chatbot inquiry uses study authority to seek information from seven companies about testing, monitoring, monetization, disclosures, age restrictions, mitigation, and personal data. The detailed order template shows the operational questions companies may need to answer. The inquiry itself is not a finding of harm or wrongdoing.
Build the evidence path before the report is due
Teams making relationship-oriented chatbots can work through the statute in this order:
- Document why the product is or is not a companion under the statutory definition. Recheck that decision when memory, personas, or engagement mechanics change.
- Keep the general human-status disclosure separate from the break and content controls for known minors.
- Assign the self-harm protocol an owner, version history, test suite, escalation route, and failure-review process.
- Record intervention events and policy versions with data minimization. Full sensitive transcripts should not be the default evidence store.
- Define the public protocol description and 2027 report fields while engineers can still change the event model.
The governor's summary of laws taking effect in 2026 describes SB 243 as a safeguard for minors. The text reaches further in one respect: the self-harm protocol applies before a companion engages with users, while the three-hour reminder applies specifically to known minors.
The three-hour notice is the visible checkpoint. Whether the protocol ran, which version ran, and what happened next are the records a product team will need when someone asks if the safeguard worked.
