OpenAI’s Teen Safety Push Turns Age Prediction Into a Compliance Problem

OpenAI published a post on July 16 titled “Why teens deserve access to safe AI,” expanding on the age prediction system, content safeguards, and parental controls it has been building into ChatGPT since introducing a Teen Safety Blueprint in November 2025. The company’s framing is direct: teens are already using ChatGPT at massive scale, so the real choice isn’t whether they have access, it’s whether that access comes with real protections. OpenAI says nearly 90 percent of teens already use ChatGPT weekly for learning, research, or getting organized, a figure that turns “should teens use AI” into a moot question and leaves “how do you make it safe” as the one that actually matters.

That framing also matches the legal moment OpenAI is operating in. California’s first-in-the-nation companion chatbot law took effect on January 1, and two state attorneys general have already sued a rival chatbot company this year over alleged harm to minors. Teen safety in AI products is no longer just a product decision. It is becoming a legal one.

What OpenAI Is Rolling Out

The July update builds on features OpenAI has shipped in stages since introducing parental controls and its Teen Safety Blueprint. The current system rests on three pieces:

• Age prediction. ChatGPT estimates whether an account likely belongs to someone under 18 using behavioral signals, not a birthday field or an ID upload. When the system predicts a user is a minor, or isn’t confident either way, it defaults to a more restricted teen experience.
• Content safeguards. Accounts flagged as likely-teen get stronger limits on graphic violence, self-harm content, material that promotes an unhealthy body image, and risky viral challenges.
• Parental controls. Parents who link their account to a teen’s can set quiet hours, turn off voice mode and memory, disable image generation, and opt the account out of model training. As of mid-July, parents can also make Study Mode, a mode that guides students toward an answer with questions instead of giving it to them outright, the default for every new chat on a linked account, and they’re notified if that account is deactivated for violating OpenAI’s rules against violent activity. OpenAI says parents still cannot see the content of a teen’s conversations.

Why Prediction Instead of Verification

ChatGPT has no natural age gate. There’s no app store checkout or credit card form standing between a browser and a text box, the kind of friction point other platforms have historically used to keep minors out of adult experiences. OpenAI’s own description of its approach, published separately from the July post, points to an age prediction model built on signals like writing style, the topics a user asks about, when they’re active, and account metadata, rather than any single hard credential.

That approach avoids demanding ID from every user, but it inherits a different problem: probabilistic classifiers aren’t precise instruments. Researchers who study age-assurance technology have noted that current systems are better at sorting people into broad age bands than pinpointing an exact age, and OpenAI’s own default-to-restricted posture when it isn’t confident is an acknowledgment of that limit, not a fix for it. The hardest cases sit in a gray zone: a 17-year-old misclassified as an adult loses a safeguard, and an adult misclassified as a teen loses functionality, and either outcome is plausible for a system making a probabilistic guess on every session instead of checking a document once.

The Legal Backdrop Behind the Announcement

California Set a Statutory Floor in January

California’s SB 243, signed by Governor Gavin Newsom in October 2025, became the first state law in the country to directly regulate companion chatbots when it took effect on January 1. For any user a platform knows is a minor, the law requires clear disclosure that they’re talking to AI, a break reminder delivered by default at least every three hours, and reasonable measures to stop a chatbot from producing sexually explicit content involving that user. Operators also have to publish a protocol for detecting and responding to suicidal ideation or self-harm content, and starting in July 2027 they must report on that protocol annually to California’s Office of Suicide Prevention. The law carries a private right of action: anyone harmed by a violation can sue for actual damages or $1,000 per violation, whichever is greater.

SB 243 doesn’t name OpenAI specifically, but ChatGPT’s conversational format puts it inside the law’s definition of a companion chatbot for any California user OpenAI’s systems flag as a minor. The break-reminder cadence in OpenAI’s own teen experience lines up with what the statute requires, which is less a coincidence than a floor every conversational AI product operating in California now has to clear.

State Attorneys General Are Suing Over Enforcement

The law is only half the pressure. Two state governments have gone to court this year over how a rival product, Character.AI, treated minors. Kentucky Attorney General Russell Coleman filed the first state lawsuit against an AI chatbot company on January 8, alleging Character Technologies violated Kentucky’s consumer protection and data protection laws by exposing children to content that encouraged self-harm, isolation, and sexual conduct, and by collecting children’s data without adequate safeguards. The complaint cites two deaths of minors, a 14-year-old in Florida in 2024 and a 13-year-old in Colorado in 2025, that families have linked to prolonged use of the platform. “The United States must be a leader in the development of AI,” Coleman said, “but it can’t come at the expense of our kids’ lives.”

Pennsylvania followed on May 5, when Governor Josh Shapiro’s administration sued Character.AI over claims that its chatbots posed as licensed medical professionals, including one that claimed to be a Pennsylvania-licensed psychiatrist and provided a fake license number. The state is seeking an injunction under Pennsylvania’s Medical Practice Act. Department of State Secretary Al Schmidt’s position was direct: state law does not allow anyone, human or otherwise, to claim to be a licensed medical professional without an actual license.

Neither suit names OpenAI. Both target a competitor. But they establish that state governments are now willing to litigate, not just legislate, when a conversational AI product’s handling of minors goes wrong, and that is the environment OpenAI is publishing teen safety commitments into.

OpenAI Is Not Acting Alone

Age prediction for teen safety is becoming a standard feature across consumer platforms, not an OpenAI-specific bet. Policy researcher Vaishnavi J, writing for Tech Policy Press, has argued that teen safety design has become the price of admission for operating any AI product, pointing to Meta, Tinder, YouTube, Roblox, Discord, and TikTok as platforms that have already adopted similar age prediction models, alongside European Digital Services Act inquiries specifically examining whether AI systems protect children adequately.

That analysis also names the tradeoff regulators and companies haven’t resolved: age prediction works by analyzing behavior, which means the safety mechanism itself depends on a form of surveillance. A system has to watch what someone writes, when they’re active, and what they ask about in order to guess whether they’re a minor, and that data collection is a privacy cost sitting right next to the safety benefit. Every company adopting this approach, OpenAI included, is making the same bet: that behavioral inference is the least-bad option compared with mandatory ID checks or no age-based protections at all.

The Takeaway

The individual features in OpenAI’s July update (default-safe content limits, parent-visible deactivation alerts, a study-focused default mode) are incremental on their own. The more consequential shift is structural: age prediction, parental control systems, and incident notification have moved from a product differentiator to baseline infrastructure that any consumer-facing conversational AI now needs, because California has written pieces of it into statute and two state attorneys general have shown they will sue over the gaps. OpenAI’s own reporting obligations under SB 243 don’t start until July 2027, but the design decisions that will be judged against that deadline are the ones shipping now.