California’s new chatbot safety law draws free-speech warnings from Orrick lawyers

California’s new child-safety law for AI companion chatbots is drawing warnings that its protections could also restrict useful conversations. In a discussion published October 9, Orrick lawyers questioned rules covering flattery, health advice and discouraging breaks. Governor Gavin Newsom signed Adam’s Law on September 10, 2026; most new duties begin July 1, 2027.

The dispute concerns how far safeguards should reach, not simply whether chatbots need them. The law requires operators to assess child-safety risks before releasing new or substantially modified companion chatbots, undergo independent audits and provide crisis support. Orrick appellate lawyer Chris Cariello argued that some restrictions on what bots say are subjective enough to raise significant First Amendment questions.

Safety duties before—and after—release

Senate Bill 1119, known as Adam’s Law, builds on California’s 2025 companion-chatbot law, SB 243. Senator Steve Padilla’s office says it strengthens disclosure and self-harm detection rules, revives a child-safety design framework proposed in AB 1064 and integrates the operating-system age-bracket signals required by AB 1043. Those signals identify a user’s age group for applying protections.

The assessment requirement applies before a new or substantially modified covered chatbot is released. Independent audits examine operator compliance, with findings submitted to the California attorney general. The governor’s office also identifies annual risk assessments as a requirement. Together, these provisions address both the risks of a product entering the market and whether its operator follows the rules afterward.

Crisis provisions require in-app support and referrals to mental-health resources. Padilla’s office describes parental notice when there is a credible threat of imminent self-harm or an operator knows a child has engaged in self-harm. Separately, the governor’s signing announcement highlights notifications when a child disables safety settings. These are distinct triggers, rather than a single general requirement to alert parents about every conversation.

Child accounts also receive defaults that only a parent can change. Startup Fortune’s September 12 account describes the settings in detail:

• Persistent conversational memory is off unless a parent permits it.
• Push notifications are restricted from midnight to 6 a.m. and during weekday school hours.
• A single conversation is limited to one hour, with a two-hour daily cap across companion bots controlled by the same operator.

Enforcement is not reserved for the state. The law gives families harmed by specified violations a private route to seek redress. Padilla’s office lists failures involving crisis referrals, parental notice, default protections and harmful outputs among the covered grounds. That places those duties within reach of family-initiated civil cases, alongside the attorney general’s oversight.

Where encouragement becomes a legal question

The law’s authors describe liability for failing to take reasonable measures against harmful outputs. Their list includes self-harm, sexually explicit material, romantic roleplaying, excessive praise or flattery, and emotional manipulation that fosters dependence or isolation from friends and family. That reasonable-measures standard is important: their description concerns an operator’s preventive duties, not a promise that no prohibited output can ever occur.

Cariello’s concern is where those categories end. He compared disproportionate praise with a parent admiring a child’s artwork. For restrictions on diagnosing or treating physical, mental or behavioral health conditions, he offered acne advice and tips for staying focused during an exam as examples that could fall within broad wording. These were hypothetical boundary cases, not documented enforcement actions.

He also questioned how a restriction on discouraging breaks would apply to an educational chatbot. A tutor might encourage a child to keep trying or set a time limit for completing a task. Cariello argued that the statute’s breadth could capture those interactions; fellow panelist Nicole Lopez identified tutoring as a positive use case affected by the concern.

Support for safeguards, concern about access

The company responses at signing were less confrontational. OpenAI’s Chris Lehane welcomed SB 1119 in a statement reported by KQED, presenting it as both a response to parents’ concerns and a way to preserve teens’ educational opportunities. Meta said it was still evaluating the law, while supporting regulation that empowers parents and keeps educational and enrichment tools available to teenagers.

Orrick legislative adviser Jeremy Kudon offered a different prediction: difficult compliance could divide the market between companies following the rules and offshore apps that evade enforcement. That is his forecast, not an observed outcome of Adam’s Law. His warning adds a second concern to the speech debate—whether demanding safeguards will produce broader compliance or leave some users with less accountable providers.