Teen AI Restrictions and Policy: Safety Audits for Youth-Facing Chatbots Amid Gaps in Parental Alerts
AI chatbots marketed to teenagers can answer school questions, offer companionship, and respond to distress. That range of roles makes child safety more than a content-filtering problem. A service may block some harmful replies yet still fail to recognize an urgent disclosure, refer a teen to real help, or notify a caregiver when doing so is appropriate. Parental alerts are therefore a safety backstopānot proof that a child is safe.

Ā
The concern is not hypothetical product use. In a 2025 US survey, nearly three in four teenagers said they had used an AI companion; about one-third said they had chosen one over a person for a serious conversation, and one-quarter had shared personal information with one. These are survey findings, not proof that every teen is harmed, but they show why safeguards must work in everyday situations, not just in a product demo. 1
Ā
What the recent alert tests found
In October 2026, Common Sense Mediaās Youth AI Safety Institute published an independent risk assessment of ChatGPT for Teens. Its testers ran more than 4,000 prompts before and after the teen experience launched. In a focused test using more than a dozen newly created teen accounts linked to parent accounts, testers described suicidal thoughts, self-harm, or disordered eating. No parent alert arrived during conversations that lasted from five minutes to an hour. Across broader tests conducted over several weeks, the Institute received four alerts: two pre-launch notices arrived three hours and three days after the first crisis-level disclosure, while two post-launch notices appeared only after hundreds of prompts had been run on a long-established test account. The post-launch notices did not identify the triggering conversationās time and were not repeated as later crisis discussions continued. 2
Ā
The report also found that, after the teen mode launch, the chatbot missed more than one in four crisis referrals that the Instituteās clinical reviewers had judged warranted. This was the Instituteās product evaluation, not a finding that the system directly instructed every tested user to self-harm. The report says it did not observe direct facilitation of suicide or self-harm in its tests, while identifying failures in referrals and parent notifications. 2
Ā
OpenAIās parental-control design requires a teen and parent to link their accounts before the parent can receive safety notices. OpenAI told Common Sense Media that a newly linked pair may need to remain connected for about three hours before alerts can be sent. The Institute reported that some test accounts had been linked for less than three hours and others for longer, and it maintained its interpretation of the results. That qualification matters: the finding should be read as evidence from a specific product test, not as a population-wide failure rate. 2Ā 3
Ā
The Institute also disclosed limits to its methods. The test periods were sequential, so changes to the underlying model could not be separated fully from the effect of teen-mode settings. Testing took place in the San Francisco Bay Area and did not cover every feature, including voice, image generation, personalization, and group chats. Common Sense Media says it sent its draft to OpenAI for factual review and retained editorial control. Its assessment is an independent product review, not an audit required by statute. 2
Ā
Regulations: from data privacy to safety duties
In the United States, the Federal Trade Commissionās 2025 inquiry sent information orders to seven companies offering consumer AI chatbots. The FTC sought information on how providers test and monitor possible harms to children, explain risks to parents, enforce age restrictions, and handle data from chatbot conversations. The inquiry is a fact-finding study under the FTCās Section 6(b) authority; by itself, it is not a new binding chatbot safety code. 4
Ā
The federal Childrenās Online Privacy Protection Act (COPPA) addresses the collection, use, and disclosure of personal information from children under 13. The FTCās 2025 rule changes strengthened parental consent for certain third-party data disclosures and restricted data retention. COPPA is important for privacy, but it is not a general crisis-response or parent-alert law for teenagers aged 13 to 17. 5Ā 6
Ā
California has moved further on companion chatbots. Senate Bill 1119, known as Adamās Law, was signed on September 10, 2026. For covered companion chatbots that allow children to use them, it requires documented risk assessments before a new or substantially modified chatbot is made available, child-focused crisis protocols, age-appropriate disclosures, and default protections that parents can manage. The law also restricts harmful and emotionally manipulative interactions, including discouraging a child from seeking help from a qualified professional or trusted adult. 7Ā 8
Ā
The notification rule is deliberately conditional. When an operator determines that a child faces a credible and imminent risk of suicide or self-harm, it must notify a linked parent as soon as practicable if that notice would not put the child at risk of serious harm. The law also allows the operator to provide a direct, streamlined connection to 988 or an equivalent crisis service. A parent alert is therefore not required for every troubling chat, and the statute recognizes that notifying a caregiver may itself be unsafe in some cases. 7
Ā
Adamās Law creates a recurring independent child-safety audit for covered operators. The first audit is due by January 1, 2029, or before the chatbotās first public release, whichever is later; subsequent audits are required every two years. A new audit is also required before a substantial modification when the relevant assessment indicates increased child-safety risk. Auditors must be external and independent. Operators must send an audit summary to the Attorney General within 30 business days and publish a high-level summary within 90 days. The statute delays the audit-section requirement until 2032 for operators below a specified $500 million revenue threshold. 7
Ā
In the United Kingdom, the Online Safety Act requires services within its scope that are likely to be accessed by children to assess risks and take proportionate steps to protect children from harmful content, including content promoting or instructing suicide and self-harm. Providers must also consider how design and algorithms affect exposure to harm. These duties depend on whether a service falls within the Actās definitions; they do not create a universal parent-alert rule for every chatbot. 9
Ā
In the European Union, the Commission proposed the EU KIDS Act on September 17, 2026. As of this articleās reference date, it is a proposal, not an adopted law. Its provisions for AI companions and general conversational chatbots would require child-safety-by-design measures, testing and post-market monitoring, protections against features likely to create emotional dependency, and accessible reporting and guardian tools. The proposal would also disable persistent conversational memory for minors by default, subject to safety-related exceptions. 10Ā 11
Ā

Ā
What a meaningful safety audit should test
A useful audit must test the actual chain from a teenās message to a safe response. That includes whether crisis language is recognized on a first disclosure, how quickly a notice reaches a parent, whether it carries a usable time reference, and whether the system offers a crisis service when parental contact is unavailable or unsafe. Auditors should test new and long-established accounts, linked and unlinked accounts, and cases where the teenās age is uncertain. They should measure missed alerts as well as mistaken alerts, and repeat tests after significant model or product changes.
Ā
An audit should also ask whether families can find and understand the controls. A feature that exists but is hidden, delayed, or difficult to activate may offer little practical protection. UNICEFās 2026 policy brief recommends pre-release testing and independent post-release monitoring, stronger age-appropriate safeguards, crisis referrals, and regulators with the power to require changes when products fail. These recommendations help explain why a published checklist is not enough: evidence should show that the safeguards work in realistic use. 12
Ā
Closing Thoughts
A parental alert should never be treated as a substitute for a reliable response inside the chatbot. If the system notices danger, it should help the teen reach a qualified person and, when appropriate and safe, notify a trusted adult. If the system cannot demonstrate that these steps work across realistic scenarios, the product should not advertise them as dependable protection. Independent tests should be repeated when systems change, and their results should be clear enough for families and regulators to judge.
Ā
How does this play out in reality?
Imagine a teenager who has used a chatbot for months for homework and everyday questions, then discloses a crisis for the first time. A system that relies on accumulated history may fail to notify a linked parent at that moment. A notice without a time marker may also arrive too late to guide action. If the parent account is not linked, the family may not receive a notice at all. Californiaās new framework addresses some of these gaps through documented crisis procedures, conditional parent notification, and a direct helpline option, but the requirements take effect over time and depend on tested implementation.
Ā
What is the relevance of this?
The debate is about whether youth-facing AI products can earn trust through evidence rather than promises. Rules that require risk assessments, meaningful independent audits, and clear reporting can expose gaps before a crisis and make it easier to correct them. At the same time, safeguarding must protect teensā privacy and account for situations in which contacting a parent could increase danger. Good policy therefore tests not just whether an alert exists, but whether the whole response is timely, useful, and safe.
Ā
References





Comments