15 SEP 2026 — California will require companion chatbots to pass independent child-safety audits and to produce annual risk assessments, under a package of thirteen bills Governor Gavin Newsom signed on 10 September. The same package bans autoplay and history-based algorithmic feeds for users under 16 and raises the civil penalties companies face when their products injure children.
Plenty of jurisdictions have passed rules about children online this year. California's audit requirement goes further. It moves the question of whether a chatbot is safe away from the company that built it and gives it to someone outside.
What was signed
The companion-chatbot bill is SB 1119, from Senator Steve Padilla. It requires crisis protocols for users who express suicidal thoughts, parental controls, and notifications to parents when a child switches off safety settings. It also requires independent child-safety audits and annual risk assessments, which the governor's office describes as the first such requirement in the country. CBS News reported that the rules also oblige companies to alert parents when young users show signs of emotional distress.
Several other bills sit alongside it. Assemblymember Josh Lowenthal's AB 2 creates civil penalties for injuries to children; CBS reported fines of up to $1 million per child where a company is found negligent. His AB 1709 covers age restrictions and establishes an e-Safety Advisory Commission. Senator Susan Rubio's SB 1276 extends child sexual exploitation law to AI-generated and digitally altered images of minors. Assemblymember Buffy Wicks's AB 1856 deals with age-verification signals for apps.
The addictive-feeds provision prohibits autoplay and algorithmic feeds built from a user's history for anyone under 16. Newsom framed that as the heart of the package: "This is about the features themselves. This is about actually addressing the problem, the scrolling algorithm."
Why the audit requirement matters
Until now, chatbot safety rules have mostly relied on the company's own account of what it does. A provider publishes a safety policy, describes its safeguards and, when something goes wrong, explains how the safeguard was supposed to work. No one outside the company tests that claim before a child encounters the product.
An independent audit brings in a party with access, a mandate and, in principle, the independence to say that a crisis protocol does not trigger reliably or that a parental control can be switched off by the child it is meant to supervise. Financial statements, safety-critical engineering and medical devices are checked this way. Software has largely certified itself until now.
What the audit requirement does not yet say
The signing announcement leaves the details open. It does not say who qualifies as an auditor, what standard an audit measures against, how often audits must run beyond the annual risk assessment, or when the obligations take effect.
California has been here before. Its separate AI auditor registry law, which we reported last week, creates a register of independent auditors but binds nobody to use one until 2029. How strong this one proves depends on a definition of the auditor that California has not yet written.
Industry objections are already on the record. Meta told CBS News that delivering a tailored experience is an important part of what makes Facebook and Instagram valuable for teenagers. That is the business case against the feeds provision, and a preview of the arguments likely to follow.
A different approach from this region's
Southeast Asia is regulating the same risk from the other end. Indonesia began enforcing restrictions on social media accounts for under-16s in March, classifying eight platforms as high-risk; Malaysia is building age verification around national identity. Both approaches decide who may hold an account.
California has mostly left accounts alone and gone after features instead: what a feed may do for a teenager, what a chatbot must do when a child is in distress, who checks. Account-based rules are easy to state and hard to enforce, since children lie about their age. Feature-based rules are harder to write but, if the auditing works, easier to verify. Which works better is still open. The two regions are now running the experiment side by side.
What to watch
Whether the audit requirement bites will come down to the regulations that define an independent auditor: who is allowed to perform audits, and whether auditors must be paid by someone other than the company being audited.
The deepfake provisions reach further than the headlines suggest. Alongside SB 1276's extension of child exploitation law to AI-generated images, the package allows civil actions of up to $250,000 against those who facilitate nonconsensual sexually explicit material, which reaches the platforms and tools that distribute it as well as the people who create it.
Legal challenges are likely, particularly to the feeds provision, which technology companies have contested on free-speech grounds in other states. A court ruling on that part of the package would not affect the chatbot audits directly, but it would shape how far California can regulate product design.
The companies themselves may move first. A chatbot provider that commissions and publishes an independent child-safety audit before the rules require it would set the benchmark everyone else is then measured against.
View author profile → · Editorial policy
