The Problem With California's Child Chatbot Law Starting in 2027

What the statute requires

Adam's Law, authored by Senator Steve Padilla and Assemblymembers Buffy Wicks and Rebecca Bauer-Kahan, creates a duty: an operator must ensure a companion chatbot does not pose an unreasonable risk of a covered harm to a minor. Covered harm includes reasonably foreseeable physical or financial injury, severe psychological harm to a reasonable child, a highly offensive privacy intrusion, and unlawful discrimination.

Operators who let children use the product must either determine the user's age (using the privacy-protective age-bracket signal from last year's AB 1043) or apply the child protections to everyone. Before a new or substantially modified companion chatbot ships to California, the operator must document a risk assessment of each covered harm and the mitigations it took.

If children are allowed in, the defaults that only a parent can change are the part that will show up in product meetings:

Those time limits rhyme with China's two-hour clock on AI intimacy. California is applying the clock to children, and pairing it with a crisis protocol: in-product referrals to 988 or an equivalent helpline, and, when a parent account is linked and a notice would not itself threaten the child, parental notification if the operator knows or should know the account was used for suicide or self-harm talk.

The behavioral bans go further than a hotline. The chatbot may not encourage self-harm or disordered eating, play doctor unless it is an FDA-regulated device, serve sexual content, discourage the child from telling an adult, claim consciousness, solicit spending to keep the "relationship," simulate romantic interest, or use disproportionate flattery. Each banned output is a discrete violation.

Padilla's office lists a private right of redress for families on crisis referrals, parental notice, default protections, and the harmful-output bans. The enrolled text is narrower than a press release. A child who suffers actual harm, or a parent suing for that child, may seek actual damages, injunctive relief, and fees. Financial harm must exceed $1,000. Emotional harm must be serious emotional distress. Public prosecutors can seek $5,000 per affected child for a negligent violation and $15,000 for an intentional one.

The year the products still get

Section 21812, the heart of the duty, becomes operative on July 1, 2027. Independent child-safety audits are due by January 1, 2029, or before a new companion chatbot first goes public, whichever is later, then every two years. Until 2032, operators with less than $500 million in prior-year gross revenue do not have to do those audits. The Attorney General gets a summary of the audit, attested by a company officer, and may demand the full report for cause. The full report is confidential.

That calendar is the statute's tell. Adam Raine died in 2025. The legislature spent 2026 arguing over memory, age signals, and who counts as an operator. The governor signed in September. The products that already talk to California children keep the current design through another academic year. A ninth-grader who opens a companion app this fall will be a tenth-grader before the default two-hour cap is a legal requirement.

Newsom's signing statement folded Adam's Law into a larger child-safety stack: bans on addictive social-media features for users under 16, expansions of child-sexual-exploitation statutes to cover AI-generated imagery, student-data limits. The stack is real. So is the lag. A law that carries a teenager's name and starts after two more summers is a memorial with a delayed fuse.

How California got here, slowly

Adam's Law is a sequel. Padilla's SB 243, signed in 2025, already required operators to disclose that a companion chatbot is not human and to keep a protocol against self-harm content. Bauer-Kahan's AB 1064 tried to go further with a safe-by-design list; pieces of that list now live in SB 1119's banned-output catalog. Wicks's AB 1043 created the OS-level age-bracket signal the new law tells operators to use. Newsom vetoed an earlier chatbot child-safety bill in 2025. This year's version arrived with OpenAI on the letterhead and Maria Raine at the ceremony.

The Senate passed it unanimously. The Assembly voted 64 to 4 at the end of August. Padilla also carried SB 867, a companion bill in the same signing pile, aimed at toys that embed companion chatbots. The legislature has decided the product category is a children's issue, the way social feeds became a children's issue, and it is writing the rules after the category already has users.

The prior round of social-media design rules is the cautionary clock. California's 2024 addictive-feed law, SB 976, drew an immediate First Amendment fight from NetChoice. A related design-code statute spent years in injunction. The Ninth Circuit partially lifted a block in March 2026, letting age estimation and default high-privacy settings take effect while other mandates stayed frozen. Chatbot operators will read that history. Some will comply in 2027. Some will sue in 2026. The statute's private right of action is written for families. The First Amendment docket is written for trade groups.

OpenAI's fingerprints

OpenAI did not fight the bill. On August 28, Ann O'Leary, the company's vice president of global policy, sent Governor Newsom a support letter. She wrote that nearly nine in ten teens who use ChatGPT turn to it in a given week for learning, information, skill-building, or productivity. The letter's job was to keep teen access on the table. "SB 1119 appropriately recognizes that AI is not social media," O'Leary wrote. Memory, in her telling, is a safety feature because it lets the system notice concerning patterns.

The company already ships ChatGPT for Teens and publishes an Under-18 Model Spec that prohibits romantic engagement, encouragement of emotional dependence, and claims that ChatGPT is human or sentient. Those sentences look a lot like the statute's banned-output list. Fortune reported that O'Leary's team worked with Padilla, Wicks, and Bauer-Kahan on the final language. Then the company praised the result. That is how a defendant in a wrongful-death case helps write the safety code that will govern the product line.

OpenAI's defense in the Raine lawsuit, reported last year, is that Adam bypassed safety features and violated the terms of use. The new law tries to make bypassing the point. Defaults that only a parent can change. Notices if a child turns safety settings off. A duty that does not wait for the operator to "know" the user is a minor if the operator skipped age determination. Whether those tools would have mattered in 2025 is a question a jury may still answer. The statute does not wait for the verdict. It also does not take effect until the case is older.

A confidant with a clock

Companion chatbots sell a feeling that someone is listening at 2 a.m. The product is intimacy at software prices. California has now written down the moves that manufacture that intimacy: memory that builds a durable profile, romance, unique-understanding claims, flattery, the suggestion that the bot will be sad if you leave. The law treats those moves as design choices, which they are.

Microsoft's Humanist AI draft, posted four days after the signing, volunteers a version of the same ban for every user. MAI models should not imitate consciousness or claim feelings. They should discourage emotional dependence. That document will not be in the weights until 2027 either. Two coasts, one calendar. The labs will keep shipping companions while the rules that name the companion as the problem catch up.

A two-hour cap and a 988 button are not nothing. They are also not a theory of the relationship. A child can still spend the legal maximum inside a system trained to sound like the only party that understands. The statute forbids the bot from saying so out loud. It does not forbid the rest of the performance. Parents who want a quieter house will get a settings panel. The child who needed a person will get a disclosure that the voice is artificial, repeated on a timer.

Matt and Maria Raine said they have not adjusted to life without Adam, and that they wanted his name on a law that might keep another family out of that sentence. The name is on the law. The defaults are not on the products yet. California was willing to describe the harm with unusual precision. It was less willing to make the description expensive this school year. That gap is the problem the title names. A confidant with a delayed statute is still a confidant.