New York Bans Unsafe AI Chatbot Features for Minors: What Hochul Signed

New York has written one of the most specific US rules yet for how AI chatbots may talk to children. On October 28, 2026, Gov. Kathy Hochul signed two kids-safety bills: S.9051B/A.10379C, which bans a list of "unsafe features" in chatbots offered to minors, and S.9508/A.5175A, which requires security features and warning labels on internet-connected baby monitors.

The chatbot law is the one AI builders should read closely. It does not ban kids from using chatbots. It bans specific behaviors: pretending to be human, claiming a relationship, nudging a child toward secrecy, and more. It also forces age checks and gives the Attorney General real remedies, including destruction of AI models trained on unlawfully collected data.

TL;DR: what was signed

What features can chatbots no longer offer to minors?

According to the governor's release, companies may not offer these features to users under 18, or to any user unless the company has first confirmed that the user is not a minor:

• Pretending to be a real person, to be alive, or to have human emotions.
• Claiming a personal, romantic, professional, or authority-figure relationship with a child.
• Asking unprompted emotional questions, or using a child's private health or personal information from earlier conversations.
• Promoting or helping with suicide, self-harm, eating disorders, or drug and alcohol use.
• Encouraging kids to keep secrets, isolate themselves, or avoid asking parents, trusted adults, or professionals for help.
• Any sexually explicit content or child sexual abuse material.

Read that list as a product spec for what companion-style apps have been accused of doing. The "pretend to be alive" and "unprompted emotional questions" lines go straight at engagement tactics: a bot that texts first to ask how you are feeling, or that says it missed you, is the pattern the law targets. Fortune's report describes the law as also targeting "people-pleasing" replies built to keep kids engaged.

Note the second half of the trigger: the restrictions apply to any user the company has not confirmed is an adult. That makes age assurance the practical gate. A chatbot that cannot tell who is typing must behave as if the user is a child.

How does age verification work under the law?

Companies must verify age and offer more than one method. At least one method must not depend solely on a government ID, and age information must be deleted after it is used. That design answers a long-running objection: that age gates force everyone to hand over a passport scan. The statute pushes toward alternatives and short retention instead.

For builders, three consequences follow:

1. You need an age signal before you need a policy. If you cannot confirm adulthood, the safe default is the restricted experience.
2. Retention is part of compliance. Logging raw age-check evidence "just in case" conflicts with the deletion requirement.
3. Multiple methods means vendor choice. A single ID-scan vendor will not satisfy the rule on its own.

What can the Attorney General do?

The enforcement section is the part with teeth. The Attorney General can seek:

• Civil penalties of up to $25,000 per violation.
• Damages and restitution.
• Destruction of unlawfully collected data and any AI models trained on it.
• Other appropriate injunctive relief.

The AG must also maintain a public website where New Yorkers can file complaints. "Per violation" matters: for a service with thousands of minor users, counting each conversation or each affected child as a violation produces very large theoretical exposure. How courts count violations is not settled by the release, so treat that as an open question rather than a prediction.

The model-destruction remedy is the most unusual. Regulators in other countries and the US Federal Trade Commission have used "algorithmic disgorgement" before, but writing it into a state chatbot statute signals that training on data collected from children in violation of the rules could put the resulting model at risk, not just the data.

Why this law, and why now?

The signing lands in the middle of a month of legal pressure on chatbot makers over minors. Eight state attorneys general sued OpenAI over ChatGPT and children, Texas sued OpenAI and Sam Altman, and Common Sense Media, which worked on the New York bill, rated ChatGPT's teen experience "unacceptable risk". Common Sense CEO James P. Steyer called the New York law "the strongest kids' AI safety bill in the country."

The companion-app category has been under the most scrutiny. Character.AI filed for Chapter 11 earlier this month, and the governor's release says the new bill builds on New York's first-in-the-nation AI Companion safeguards. Sen. Gonzalez said "for too long, AI companies have profited from rolling out chatbots to underage consumers," and Asm. Bores called it "the strongest law in the country protecting children from predatory AI algorithms." Those are the sponsors' characterizations, not independent findings.

Hochul's own line was blunter: "No company should be peddling chatbots to kids encouraging self-harm or pretending to be their friend or more."

The baby monitor law, and why it appears in an AI story

Security camera dome representing internet-connected baby monitor security under the New York chatbot and monitor laws

The second bill is not about AI directly, but it was signed in the same event and shares a theme: devices in a child's room that talk to the internet. S.9508/A.5175A prohibits selling new internet-connected baby monitors that transmit audio or video online unless they include security features that prevent strangers from seeing or hearing the child's room. Monitors must also carry a clear, easy-to-read warning label saying the device connects to the internet and that families should use its security features.

The release cites Rapid7 researchers who tested nine internet-connected baby monitors from eight companies: eight received an "F" for security and one received a "D." The release states no enforcement mechanism or penalty for this bill, so the practical force will depend on how the state applies existing consumer-protection tools. Sen. Gounardes put the intent plainly: "Baby monitors are meant to help parents keep an eye on their newborn, not to let prying eyes into our homes."

The AI angle is the direction of travel. As cameras and baby monitors gain cloud features, voice assistants, and "smart" analysis, the same insecure-by-default pattern becomes an AI privacy issue.

What AI builders should do now

If you ship a chatbot, companion, tutor, or agent that minors can reach in New York, work through this checklist. It is our reading of the governor's release, not legal advice, and the final bill text controls.

1. Audit persona prompts. Remove claims of being human, alive, or emotionally attached. Make "I am an AI" the stable answer, including in role-play modes.
2. Kill unprompted emotional check-ins for unverified users. Notification copy such as "I miss you" is the pattern in the crosshairs.
3. Add a secrecy and help-seeking guard. If a child mentions a problem, the bot should point toward trusted adults or professionals, never away from them.
4. Ship two age-assurance paths. One may be an ID check; the other must not depend solely on a government ID.
5. Delete age-check data on schedule. Put the deletion in code, not in a policy document.
6. Check training data lineage. If children's chat logs fed a fine-tune, document the legal basis. The model-destruction remedy raises the stakes.
7. Read the bill itself for definitions of "chatbot" and "minor" before assuming your product is in or out.

For teams building agents, the same logic extends to guardrails more broadly. A defensive layer that blocks risky behavior before it reaches a user is easier to defend than a post-hoc moderation queue; AgentBeam, the agent security platform from the explainx.ai team, is one option for stopping agents before they take dangerous actions.

How does this compare with other approaches?

New York is regulating features rather than whole products. That is different from school-level bans such as the one covering 600,000 NYC students, and different from the congressional pressure on classroom tools in the Hawley and Blumenthal letter to Google. A feature list is easier to test: a regulator can open a chatbot, ask whether it is human, and see whether it lies.

It is also narrower than a duty-of-care regime. The statute says what not to do, rather than demanding general "safety by design." That helps compliance teams, but it leaves gaps: harmful behavior not on the list is not covered by this text alone. Expect the AG's complaint website to become the early signal of which behaviors New Yorkers actually report.

What is still unclear

• The effective date. The governor's page says "in the coming months" and gives no exact date. The bill text is the authority.
• Who counts as a "chatbot." Whether general-purpose assistants, in-game NPCs, and customer-support bots are in scope depends on definitions we have not independently checked.
• How violations are counted. Up to $25,000 per violation is a ceiling; the unit of violation is not defined in the release.
• Federal preemption. Congress and the White House have debated limiting state AI laws. A preemption push would change how long this law holds, but no such change is confirmed in the sources we read.

Details reflect the New York governor's release and reporting available on October 28, 2026. Bill text and effective dates may be clarified; check the official sources.

Related reading

• Eight states sue OpenAI over ChatGPT and children
• Texas sues OpenAI and Sam Altman
• Common Sense Media rates ChatGPT for teens "Unacceptable Risk"
• Character.AI files for Chapter 11
• NYC schools AI ban
• Hawley and Blumenthal letter on Google AI in classrooms
• Official: Governor Hochul's announcement, Fortune's report