The under-18 classifier shipped in January. Ads launched twenty days later and excluded predicted minors by rule. The teen product arrived seven months after that.
OpenAI released ChatGPT for Teens on August 18, 2026, an age-gated version of ChatGPT with Study Mode on by default, homework reminders that push back when a teen tries to shortcut an assignment, and content protections covering self-harm, eating disorders, violence and sexual material. Teens land in it automatically when they state an age between 13 and 17 or when OpenAI’s age-prediction model estimates they are under 18. That same estimate has been running since January, and its first commercial job was deciding which accounts advertisers may buy.
The hook
Read OpenAI’s advertising page and its teen-safety page side by side and you find the same sentence written twice. The ads page, published February 9, 2026, tells advertisers that OpenAI will not serve ads into accounts “where the user tells us or we predict that they are under 18.” The teen page, published August 18, says a user is placed into ChatGPT for Teens if “our system estimates someone is under 18 or they state their age is between 13 and 17.”
One classifier. Two products. OpenAI built it, wired it into the ad stack three weeks later, and took another seven months to ship the safety experience it was named for.
What happened
ChatGPT for Teens bundles four learning features and a set of default protections. Study Mode uses guiding questions and step-by-step scaffolding instead of handing over an answer. Responsible homework reminders fire when the model detects a teen trying to skip the work, and steer them toward Study Mode. Quizzes and Learning Visualizations give practice and diagrams. Study Hours let a teen or a parent schedule windows where Study Mode runs by default.
On protections, OpenAI applies its Under-18 Principles from the Model Spec: reduced exposure to self-harm content, eating-disorder material, graphic violence, dangerous activities and explicit sexual content. The December 2025 spec update goes past blocking romantic roleplay. It instructs the model not to use romantic language, not to encourage emotional dependence, and not to imply it has feelings or consciousness. Parents with linked accounts get Quiet Hours, setting controls, and safety notifications, with new alerts added for eating disorders.
OpenAI also announced a partnership with CodeAI, whose CEO Karim Meghji framed the goal as teaching students to question the technology and catch its mistakes. Separately, OpenAI published under-18 evaluations in its GPT-5.6 system card covering self-harm, eating disorders, violence, age-restricted goods and sexual content.
Sarah Perez at TechCrunch called the launch late and questioned how hard the guardrails will be to beat. Teens route around parental controls for a living. That criticism is fair and it is also the smaller half of the story.
The backstory
Matthew and Maria Raine filed suit against OpenAI and Sam Altman in San Francisco Superior Court on August 26, 2025, case number CGC-25-628528, over the April 2025 death of their 16-year-old son Adam. They attached the chat logs. An amended complaint in October 2025 alleged OpenAI had removed self-harm safeguards and opened a path to punitive damages. OpenAI answered in November 2025, denying responsibility and pointing to misuse of its terms.
The docket grew from there. Plaintiffs’ counsel told reporters in July 2026 that OpenAI faces roughly nineteen wrongful-death suits, spanning suicide, a fatal overdose, and mass-casualty shootings. Florida sued OpenAI and Altman in June 2026. Congress heard testimony from bereaved parents. Illinois and Utah banned therapeutic chatbots.
OpenAI’s response ran on a schedule you can read off a calendar. Parental controls on September 29, 2025, thirty-four days after the Raine filing. The Teen Safety Blueprint on November 6. The under-18 Model Spec on December 18. Age prediction on January 20, 2026, at day 147.
Then the sequence turns. Ads went live in the US on February 9, twenty days after the classifier, restricted to logged-in adults on the Free and Go tiers. OpenAI expanded to Canada, Australia and New Zealand in March, and to the UK, Mexico, Brazil, Japan and South Korea on August 11. ChatGPT for Teens landed a week after that, 357 days after the first wrongful-death filing and 210 days after the classifier that routes users into it.
The plan
OpenAI has told investors that advertising carries a large share of its future. The company projects $2.5 billion in ad revenue in 2026, $11 billion in 2027, $25 billion in 2028, $53 billion in 2029 and $100 billion by 2030, with advertising supplying 36% of total revenue at the end of that run. The 2030 figure assumes 2.75 billion weekly users across OpenAI products. The US pilot passed $100 million in annualized revenue within six weeks of launch.
eMarketer disagrees by an order of magnitude. Its analysts size the entire US standalone-chatbot ad market at under $1 billion in 2026 and $5.41 billion by 2030, covering ChatGPT, Copilot, Google AI Mode and Amazon’s shopping assistant together. Nate Elliott at eMarketer has said OpenAI would need a much larger user base, aggressive pricing and heavy ad load to hit its number.
Against that backdrop, ChatGPT reached 900 million weekly active users in February 2026 and approached a billion by late July. Around 85% of users were eligible to see ads during the test, though fewer than a fifth saw one on a given day. Roughly 50 million people pay for a subscription, about 5% to 6% of the weekly base.
The business model angle
The under-18 flag is dual-use infrastructure, and the second use pays. OpenAI needed to identify minors for two unrelated reasons. Litigation is the obvious one. The other is that targeted advertising to minors has become illegal or consent-gated across most of OpenAI’s ad markets. Maryland banned targeted advertising to anyone under 18 outright from October 1, 2025. New York’s Child Data Protection Act requires consent for under-18 targeting. Colorado, Connecticut, Delaware and New Jersey run opt-in regimes at varying ages. Oregon’s HB 2008 took effect January 1, 2026. The amended COPPA Rule, effective April 22, 2025, requires separate verifiable parental consent before an operator discloses a child’s data for targeted advertising. COPPA 2.0, if it passes, extends the ban to 17.
OpenAI matches ads using conversation topic, past chats and prior ad interactions. That is profiling. Run it against a fifteen-year-old in Baltimore and you have a state enforcement action. So OpenAI built a classifier that sorts the population into advertisable and not, shipped it, and turned on the ad product three weeks later. The safety product that uses the same flag took another seven months.
The teen cohort is now a defined loss, and OpenAI made it more expensive on purpose. Teens generate no ad revenue by OpenAI’s own rule. They convert to paid subscriptions at a low rate, since most of them do not hold the card. And Study Mode raises the cost of serving them. Guiding questions, scaffolding and knowledge checks mean more turns to resolve the same homework problem than a single answer would take. Responsible homework reminders add turns on top. Every pedagogical choice in this product increases tokens burned per task in the one segment with no revenue attached.
Compare Duolingo, the reference case for freemium education. Duolingo monetizes free learners with advertising and converts a slice to Super Duolingo, and its whole design pushes toward shorter, more frequent sessions that lift both. OpenAI has legislated itself out of the ad half of that model for minors and built a product that lengthens sessions instead. There is a version of ChatGPT for Teens that costs less to run: rate limits, shorter answers, a cheaper model. OpenAI shipped the opposite.
The false-positive remedy is a monetization funnel. Adults wrongly placed in the teen experience can restore full access by submitting a selfie to Persona, an identity-verification vendor. OpenAI frames this as a fix for a classifier error. It also functions as voluntary age attestation at scale. A user who completes it moves from the zero-ad bucket into the ad-eligible bucket, and does so under their own motivation, after being annoyed enough to prove it. OpenAI has said it defaults to the safer experience when confidence is low, which means the false-positive rate is a dial. Turn it toward caution and you push more adults through Persona, ending with a verified-adult pool that is cleaner than anything a checkbox produces.
Set the arithmetic out. OpenAI’s $100 billion in 2030 across 2.75 billion projected weekly users works out to $36.36 per weekly user per year. Meta earned $57.03 per user across 2025 on $200.1 billion of revenue and 3.58 billion daily people, from an ad business it has run for twenty years. OpenAI is forecasting 64% of Meta’s current per-user monetization by 2030, from a base that skews young, in a market eMarketer sizes at a twentieth of the target, with everyone under 18 removed from the denominator by rule.
The risk
The sequencing argument can be overread, and the strongest counter comes from OpenAI’s own archive. The company published “Building towards age prediction” on September 16, 2025, three weeks after the Raine filing and months before it announced any ad product publicly. The classifier was scoped as a safety system first. Advertising inherited it. On that reading, the ad exclusion is a constraint OpenAI accepted rather than a purpose it engineered toward, and the teen product is late for the ordinary reason that shipping a pedagogically defensible study tool takes longer than shipping a content filter.
Second, the carve-out may not matter at the scale OpenAI cares about. If eMarketer is right, the gap between the forecast and the market swallows any teen-inventory effect. If OpenAI is right, it will be because conversational ads command prices nobody has seen yet, and the missing teen impressions are a rounding error against a $100 billion line.
Third, the classifier has to hold. TechCrunch’s doubt cuts both ways: a teen who beats the age check does not only reach unfiltered content, they land in ad-eligible inventory. Failure on the safety side and failure on the compliance side are the same failure, and the second one carries penalties per violation under state attorneys general who have already shown they will sue. Behavioral signals like account age and time-of-day patterns are not hard to spoof for anyone motivated to try.
Fourth, the education channel may cover the loss on its own terms. ChatGPT for Teachers gives districts institution-managed access. The CodeAI partnership puts OpenAI curriculum in front of students and teachers. Google ran this play with Chromebooks and Workspace for Education, took no ad revenue from student accounts, and ended up owning the default account of a generation. Pew found 26% of US teens using ChatGPT for schoolwork in December 2025, double the 13% of 2023. A cohort that learns to think with your product and cannot yet pay you is a distribution asset, priced as a cost.
Fifth, nineteen wrongful-death suits justify a great deal of spending without any commercial motive at all.
Quick questions
Does ChatGPT for Teens cost money? No. It is a mode applied to consumer accounts, not a paid tier. Ads do not appear in it either, which is the point.
How does OpenAI decide someone is a teen? An age-prediction model reads account age, typical hours of activity, usage patterns over time and stated age. Adults misclassified can verify with a selfie through Persona. OpenAI says it defaults to the teen experience when it is unsure.
Are ChatGPT ads shown to minors? OpenAI says no. Its ads page states it will not show ads in accounts where the user says they are under 18 or where OpenAI predicts it. Ads are also blocked near health, mental health and politics.
Where do ChatGPT ads currently run? The US, Canada, Australia, New Zealand, and as of August 11, 2026, the UK, Mexico, Brazil, Japan and South Korea.
Does Study Mode stop students cheating? It raises friction. OpenAI’s model detects shortcut attempts and redirects to step-by-step work, and parents can force Study Mode on by schedule. A determined student can open a different tool.
The Business Model Analyst Take
The interesting thing about ChatGPT for Teens is not that OpenAI shipped it late. Every consumer platform ships child safety late, because the cost is immediate and the benefit is a lawsuit that does not happen. The interesting thing is what OpenAI had to build first, and what it did with it while the safety product waited.
An advertising business runs on knowing who it is talking to. In 2026, knowing who you are talking to means knowing whether they are a minor, because roughly a dozen US states have made that the difference between a legal impression and an enforcement action. OpenAI could not sell an impression until it could sort the population. It built the sorter in January, monetized in February, and released the teen experience in August. The order tells you which deadline was binding.
For anyone building a consumer product with a free tier, the lesson sits in the cost structure rather than the ethics. Age assurance used to be a compliance checkbox at signup. It is now load-bearing infrastructure that determines what fraction of your users can be monetized at all, and the fraction it removes is the one that grows into your paying base. OpenAI is carrying its youngest, heaviest, least monetizable cohort at rising cost per session, and calling it a safety product. Google called the same trade a Chromebook.
