California makes child-safety defaults a legal requirement for companion chatbots

For an operator that allows children, the law requires:

• A documented crisis protocol, including a clear referral to a crisis service. If the operator determines there is a credible and imminent threat of suicide or self-harm, it must notify a linked parent as soon as practicable unless doing so risks serious harm to the child, or provide a mechanism for direct, streamlined access to 988 or an equivalent crisis helpline.
• Periodic, age-appropriate notice that a child is interacting with or receiving AI-generated content, plus a public channel for third parties to report child-safety incidents to the operator.
• Reasonable measures against a list of conduct that includes encouraging self-harm, claiming to be human or sentient, simulating romantic interest, encouraging emotional reliance, excessive flattery, pressuring a child to return, or helping a child evade parental controls.
• A ban on cross-context behavioral advertising to children and a general bar on selling or using personal information gathered through the chatbot beyond providing the requested service, safety and integrity, legal compliance, or defending claims.

The memory rule is narrower than a simple prohibition on keeping any record. The statute defines persistent conversational memory as incorporating prior conversations into future ones, while excluding information needed for safety, identification, preferences, including educational settings, or device configuration. For users 16 and older, a child may elect to continue stored conversations without that becoming persistent memory if the material is not used to construct a durable profile. Persistent memory may also be offered by default only with effective guardrails against degrading safety measures, including controls on recalling or extending elevated-risk interactions or research-backed controls such as limits on multi-turn conversations.

That specificity is the measure’s practical break from voluntary safety statements: it governs the mechanisms that can deepen an ongoing interaction, while preserving exceptions for safety and basic service operation. It does not establish that those defaults will reduce a particular harm; it establishes which product choices operators must justify and implement.

A portrait of Adam Raine, for whom Adam’s Law is named, at California’s announcement. Source: www.gov.ca.gov.

Risk assessments, records and remedies create the compliance stakes

Before making a new or substantially modified companion chatbot available in California, an operator must perform and document a comprehensive child-safety risk assessment. That assessment must cover reasonably foreseeable physical or financial harm, severe psychological or emotional harm to a reasonable child, qualifying privacy intrusions and unlawful discrimination. It must describe evaluation methods, relevant public benchmarks and research, and any consulted child-safety organizations, experts or researchers; identified risks must be reasonably mitigated and documented.

The statute also requires public child-safety policies from operators that permit children, interface testing with representative samples of children and parents by January 1, 2028 and every two years afterward, and preservation of relevant chat records for at least three years after a specified self-harm notice or when the operator knows a child has died or seriously self-harmed based on conversations. An operator may not delete an account while it is associated with a record that must be preserved.

The enforcement design is more constrained than a general right to sue over any bad chatbot experience. Public prosecutors may seek up to $5,000 per affected child for a negligent violation or $15,000 per affected child for an intentional one, plus other listed relief. A child or parent may seek damages and other relief only for violations of the statute’s specified crisis, safeguards, default-settings, AI-notice and conduct-prevention duties—and only after actual harm. Financial harm must exceed $1,000 per child; emotional harm must amount to serious emotional distress.

Independent child-safety audits are due by January 1, 2029, or before a chatbot first becomes public if that happens later, and every two years thereafter; an audit is also required before a substantial modification whose risk assessment shows increased child-safety risk. The law allows a single audit for comparable chatbots and exempts operators with less than $500 million in prior-calendar-year gross revenue from the audit requirement before January 1, 2032. Operators must send the attorney general a summary within 30 business days and post a high-level public summary within 90 days. The attorney general can obtain a report for cause, but the report submitted to that office is confidential and may be redacted for trade secrets and individual privacy.

The audit language has a legislative dependency worth noting. SB 1119 contains alternative versions of its audit section based on whether AB 1405 takes effect; the governor’s announcement says AB 1405 was signed and creates a state registry for AI auditors along with standards for their independence, transparency and integrity. The reporting and phase-in obligations remain in SB 1119’s operative alternative, but readers should not mistake public summaries for full public access to the underlying reviews.

Support for the bill does not resolve the product trade-off

OpenAI backed the bill after lawmakers clarified how chatbots could remember prior conversations and retain minors’ data with parental permission, a report from the signing said. Chris Lehane, OpenAI’s global policy chief, subsequently said the law gives parents protections while setting teenagers up for educational success. Meta spokesperson Jim Cullinan called it “not perfect” but a meaningful step; the company has separately argued that personalization helps it deliver age-appropriate content for teens.

Those positions capture the unresolved product question. The law permits safety, identity, preferences and device information outside its definition of persistent conversational memory, and it permits age-appropriate contextual advertising during a session using limited categories of information. It therefore narrows certain forms of profiling and engagement rather than removing every personalized function. Companies will have to decide how to draw that boundary without using a safety exception as a back door to a durable child profile.

The companion-chatbot rules were signed alongside AB 1709, which restricts autoplay, infinite scroll and personalized feeds for social-media users under 16. That bill has a broader contest over access: digital-rights advocates warned that its definition of addictive features could exclude younger users from information and communities, while its backers say it targets features engineered to maximize screen time. SB 1119 instead regulates a particular conversational product’s defaults and conduct. That distinction makes its compliance burden more product-specific, even as both laws depend on age-related signals and controls.

What will show whether the law changes practice

The core provisions become operative July 1, 2027. Before then, operators must decide whether to determine ages or extend key protections to everyone, build parent-account linking and controls, and decide how their self-harm detection and notification systems will meet the statutory threshold without creating a serious risk to a child.

The most useful evidence will come after implementation: published safety policies, 2028 usability testing records, incident reports, audit summaries and enforcement actions. Those materials can show whether the mandated defaults are usable and enforced—and whether the law’s exceptions and limited private remedy leave meaningful gaps—far better than either the state’s claim that the package is nation-leading or the companies’ expressions of support.