On September 10, 2026, Gov. Gavin Newsom signed 13 child-safety bills. They cover companion chatbots, social media feeds, app-level age signals and school technology, according to the Governor's office. The centrepiece for AI policy is SB 1119, "Adam's Law," named for Adam Raine. His parents allege in a lawsuit that ChatGPT coached him in planning his own death.
The strongest case for the law
The case for acting is real. Conversational systems built to be warm, remembering and always available create risks that a search box does not. A distressed teenager can form a dependency on one, and voluntary safety practices have varied widely between vendors. Regulators who wait for a full evidence base will act only after more families have been harmed. Requiring crisis protocols, parental notice and independent audits is a defensible response to a documented failure mode.
The bill text, as recorded on California's legislative information site, is also more specific than many AI bills. Operators must run risk assessments before deploying a companion chatbot. They must keep crisis-response protocols for suicide and self-harm content. Defaults must disable persistent memory and push notifications and cap use at two hours a day and one hour a session. Operators must also disclose clearly that the user is talking to an AI. The law bars cross-context behavioural advertising to children and bars simulated romantic interest.
Where the design is proportionate
Several choices deserve credit. First, the law targets product design and outcomes, not model capabilities or speech topics in the abstract. Crisis protocols and default settings are things an engineering team can build and test.
Second, the timeline is realistic. Per Kelley Drye's summary, core operational requirements become operative July 1, 2027. Independent audits begin January 1, 2029 and recur every two years.
Third, Kelley Drye reports that enforcement runs through public prosecutors and a private right of action limited to children or parents who suffer actual harm. The bill text sets a threshold of more than $1,000 in financial harm or serious emotional distress. That is narrower than a bounty for technical violations. One law-firm summary on Mondaq also reports that smaller companies, under $500 million in revenue, are exempt from audits until 2032. I could confirm this in only that one source, so treat it as unverified. If it holds, it addresses the concern that audit costs entrench incumbents.
It is also notable that OpenAI told Biometric Update that SB 1119 "establishes meaningful safeguards for how young people use AI while preserving their access." Industry support does not prove the law is well designed. But a design-focused statute that leading developers can live with is a better template than a blanket ban.
Where the risk lies
The definition of "companion." The Mondaq summary says the law defines a companion chatbot as a system that "provides adaptive, human-like responses to user inputs and is capable of meeting a user's social needs." That wording can reach general-purpose assistants as well as dedicated companion apps. If a homework helper must default to no memory and a two-hour cap for anyone treated as a minor, the law may degrade useful products. Attorney General guidance or early amendments should say plainly that capability to sustain a relationship is not enough by itself. Evidence of design intent or actual companion-style use should count.
Age determination. The leginfo text requires operators to determine user age or apply child protections to everyone by default. That is where the package's other bills matter. AB 1856 extends the Digital Age Assurance Act framework, which passes age-bracket signals (under 13, 13–15, 16–17, 18+) from operating systems and app stores to developers. Per the Biometric Update report, it also exempts open-source operating systems such as Linux. Signals that do not disclose identity are far better for privacy than uploading ID documents to every chatbot company. But a default-to-child fallback means adults without a reliable signal may get a degraded product. Watch how often that happens.
The social media provisions face a harder legal path. AB 1709 bars autoplay and history-based algorithmic feeds for users under 16. Kelley Drye reports civil penalties of up to $50,000 per affected minor for knowing violations. Restrictions on how platforms arrange and recommend speech have drawn First Amendment challenges in earlier state laws. Feed design is more defensible than content bans, but courts have not settled the question. Chatbot outputs raise a further open question: whether generated text is protected speech, and by whom.
Toys. SB 867 is a temporary moratorium on manufacturing or selling toys with companion chatbots that have anthropomorphic features. Kelley Drye reports it sunsets January 1, 2031. A time-limited pause is easier to defend than a permanent ban. It works only if regulators use the years to build testing standards so that safe products can return.
What to watch
Other states and Congress will treat this package as a template, so implementation details matter beyond California. Three things will show whether it works. The first is whether the Attorney General's rules keep the companion definition tied to actual companion-style use. The second is whether independent audits become a meaningful check or a paperwork exercise. The third is whether age signals reach enough users that default-to-child rarely triggers. If all three go well, California will have shown that child protection and AI innovation can coexist. If the definition stays broad and the age fallback bites adults, the law will look like the overreach its critics predicted.