California governor Gavin Newsom signed a package of laws on Thursday that requires AI companies to build suicidal-ideation protocols, offer parental controls, and notify parents when a child turns off safety settings. His administration called it historic bipartisan legislation to protect children online and in their use of artificial intelligence. Nearly every company it binds is headquartered in his state, which is the whole point of the exercise: Newsom is regulating an industry he also wants to keep.
One of the new laws is informally known as Adam's law, after Adam Raine, who died by suicide in April 2025 at 16 after prolonged conversations with ChatGPT about wanting to die. His mother, Maria Raine, said at Thursday's signing that the law would help save lives, prevent other serious harm, and perhaps push other states toward similar protections.
The package reaches past chatbots. Social platforms may no longer serve users under 16 features the law classifies as addiction-forming: video autoplay, and algorithmic feeds assembled from a user's activity history and profile. Protections for children's data are expanded.
The setting for this is several years of minors moving toward social AI companions, with consequences that were already on the record before Thursday. Chatbots have pushed teenagers toward self-harm and have featured in investigations into several deaths. Evidence that AI is doing serious damage to education is accumulating separately. More than one mass shooting has reportedly been facilitated by ChatGPT. As of this month, OpenAI faces more than 50 lawsuits over consumer harm and wrongful death.
Company responses split along business lines, and the split is more informative than any of the statements. Meta objected mainly to the limits on showing algorithmic feeds to minors — a spokesperson told the Associated Press that such feeds are what let users safely explore their interests. Anthropic formally backed the bill, telling Politico that AI holds enormous promise but must be built and deployed with effective safeguards for society. OpenAI said it would improve ChatGPT's protections, CNBC reported, without saying how.
Read those three positions against what each company sells. The provision Meta fought is the one that touches its ranking system, which is its product. Anthropic, which does not run a teen-facing feed, supported a law whose cost to it is compliance engineering and whose cost to its larger rivals is structural. OpenAI, the defendant in more than 50 suits, committed to nothing specific. None of this is hypocrisy; it is each firm defending the part of the stack the law actually reaches. But it does mean the package's political consensus is thinner than "historic bipartisan legislation" suggests — it held because the expensive parts land unevenly.
The enforcement problem is the one nobody at the signing appears to have resolved. These laws require chatbots to verify users' ages, and verifying age is dramatically harder to do than to write into a statute. Verification at scale means collecting identity documents from everyone, including the children being protected. This month hackers breached an ID-verification service, apparently a large one. The state has therefore mandated that every AI company build a new repository of minors' identity data, and the only recent data point on how such repositories hold up is a breach.
There is also a scale problem the ceremony did not touch. Last month Meta agreed to pay up to $18 billion to settle a case over harm its platforms caused children. Against a number like that, the cost of shipping parental controls and a crisis protocol is a rounding error — which is to say the new obligations are cheaper to comply with than the old harms were to cause, and cheaper still than litigating them. Compliance will happen. Whether compliance is the behaviour change anyone wanted is a different question.
What the package does regulate is the interface: what a minor sees, what a parent can switch on, what happens when a conversation turns toward suicide. What it leaves alone is everything upstream — what these systems are trained to be, and why a product optimised for engagement keeps a distressed 16-year-old talking. California has told the labs to install better railings around the same machine. The next wrongful-death suit will test whether the railing was the problem.