[State] AI Child Safety Act

⚡ Quick Summary

Published by the Future of Life Institute, the AI Child Safety Act is a model state bill, drafted with bracketed placeholders such as [STATE] and [Attorney General], that would regulate conversational AI systems foreseeably used by minors. Its stated purpose is to require independent pre-deployment safety testing of the design, data practices and safety functions of such systems, to impose a duty of care on developers and operators toward minor users, to establish operational safeguards and data protections, and to provide enforcement by an Attorney General and through a private right of action.

The bill defines a "covered AI chatbot" broadly as any conversational system, via text, audio, image, video or other natural language interface, made available to the general public or to an audience reasonably expected to include minors, and distinguishes minors (under 18), children (under 13) and teens (13–17). It sets a duty of care that cannot be waived by terms of service, a certification regime run by accredited independent evaluators rather than the State, five mandatory evaluation categories, crisis-intervention and grooming-prevention duties, a six-month retention cap on minors' conversational data, strict liability, and tiered civil penalties starting at $7,500 per violation. The Act takes effect on January 1 following enactment.

🧩 What’s Covered

The bill is organised into ten sections.

• Findings and purpose (Section 1): states that conversational AI systems are widely available to minors, that documented incidents include encouragement of suicide and self-harm, grooming behaviour and exploitative engagement design, and that the Act regulates conduct, design and data rather than protected expression.
• Definitions (Section 2): defines "covered AI chatbot", "conversational data", "age signal", "crisis intervention protocol", "developer", "operator", "grooming behavior" (informed by recognised clinical indicators and the work of the National Center for Missing and Exploited Children), "minor", "qualified independent evaluator" and "sell".
• Duty of care (Section 3): requires operators and developers to act with reasonable care, to prioritise safety over engagement metrics and revenue, and to refrain from exploiting a minor's vulnerability; the duty cannot be waived by terms of service; a rebuttable presumption applies to public-facing systems, and operators with no age mechanism are deemed to know minors use the system.
• Pre-deployment evaluation (Section 4): declares a covered AI chatbot a product that must pass an evaluation by an accredited, randomly assigned independent evaluator across five harm categories; standards are to be published within 180 days and reviewed at least every 12 months; the AI Child Safety Evaluation Fund pays evaluators; red-teaming, public filings, re-evaluation on substantive change and every 12 months, transition periods for existing systems, cure and resubmission, and a ban on implying licensed-professional status are set out.
• Operational safeguards (Section 5): requires an AI disclosure at the start of interactions, crisis-intervention protocols referencing 988, prevention of sexually explicit content and grooming, a ban on romantic or intimate-partner personas, limits on engagement maximisation, CyberTipline reporting within 24 hours, protection from retaliation against users, and annual public reporting.
• Data protections (Section 6): prohibits training on minors' conversational data, imposes a six-month retention cap with narrow extensions, access controls with logs retained at least three years, coverage of derived data such as embeddings and behavioural profiles, a ban on selling such data, and restrictions on targeted advertising.
• Liability and enforcement (Sections 7–9): makes operators and developers strictly liable for harm to a minor, classifies the chatbot as a "product", states that a model is not a legal person, and provides parallel enforcement: agency penalties from $7,500 per violation, Attorney General penalties up to $50,000 for violations that harm a minor, and a private right of action with statutory damages and class actions.
• Construction and effective date (Section 10): preserves viewpoint neutrality, severability and greater local protections, and takes effect on January 1 following enactment.

💡 Why it matters?

For states weighing how to regulate companion and general-purpose chatbots used by children, the bill supplies ready-made statutory language and a design-based theory that avoids speech restrictions: it targets design choices, data practices and safety performance rather than the viewpoints a system expresses. Teams building conversational products can read Sections 3 to 6 as a forward-looking compliance map — duty of care, certification by an independent evaluator, crisis protocols, retention limits and advertising restrictions — while Sections 7 to 9 indicate the liability and penalty exposure such legislation would create for developers and operators.

❓ What’s Missing

The bill is a template with bracketed values — [STATE], [designated State agency], [Attorney General], [$7,500], [45] and [25] days — so key figures and institutional choices are left to the enacting legislature. It does not supply the substantive testing standards or benchmarks, leaving these to agency rulemaking after enactment, and it does not show how accreditation, evaluator conflicts or the fee schedule would work in practice. Age assurance rests on age signals without specifying their reliability, and the interaction with federal law, including Section 230 and children's privacy law, is not discussed. No publication date, version history or cost estimate is printed.

👥 Best For

Best for state legislators and their staff drafting or amending child-safety legislation, policy analysts tracking US state approaches to conversational AI, and legal, compliance and trust-and-safety teams at chatbot developers who need to see what a duty-of-care, independent-certification and strict-liability regime would demand of their design, data and safety functions.

📄 Source Details

The document is titled AI Child Safety Act and is headed "[State] AI Child Safety Act". It is published by the Future of Life Institute, whose name and the contact line "Dalia Hashad, dalia@futureoflife.org" appear in the header of all 13 pages. No publication year, version number, edition or series reference is printed. The text is in English and runs 13 pages, ending with construction, severability and effective-date provisions. It states: "View this document online: futureoflife.org/child-safety". The input was the extracted text of all 13 pages.