Age Prediction AI Signals: The Data Behind "Under 18"

On August 18, 2026, OpenAI launched ChatGPT for Teens and started placing accounts into it automatically. Two things put a kid there: stating an age between 13 and 17, or a model guessing they are under 18. The age prediction AI signals behind that guess are documented, and they are more mundane and more interesting than most parents assume: how long the account has existed, what time of day it is typically used, how usage patterns have evolved, and any age the user stated. No face scan. No school records. Just behavior over time, read statistically. Whether you think that is reassuring or unsettling probably depends on which part you focus on, and both reactions are defensible.

Key Takeaways

• The documented signals are account age, time-of-day activity, longitudinal usage patterns, and stated age. Topics raised in conversation are also reported as a signal.
• The system is deliberately asymmetric: when confidence is low, it defaults to the safer teen experience. False positives on adults are the intended trade-off.
• Adults misclassified as minors can restore access through a selfie-based identity check run by the vendor Persona.
• Being placed in ChatGPT for Teens changes the product: Study Mode defaults, homework reminders, no romantic language, stronger instructions against implying feelings, tighter content limits.
• Parental controls include quiet hours and study hours, and explicitly do not include reading your teen’s conversations.

What OpenAI actually shipped, and when

Age prediction rolled out on ChatGPT consumer plans starting January 2026 and became the gate for ChatGPT for Teens on August 18, 2026. The mechanics, per OpenAI’s documentation and contemporaneous reporting: if a user states an age of 13 to 17, or if the prediction model estimates under 18, the account is placed in the teen experience automatically.

The teen experience is a different product, not just a filter. Study Mode uses questions and step-by-step guidance instead of finished answers. “Responsible homework reminders” try to detect when a teen appears to be shortcutting an assignment. The model is barred from romantic language and terms of endearment, and is more strongly instructed not to suggest it has feelings, consciousness, or emotions, which OpenAI frames as reducing anthropomorphism and emotional dependence. Content limits tighten around graphic violence, viral challenges, sexual or violent role play, self-harm depictions, and content promoting extreme dieting or beauty standards.

Parental controls are worth being precise about because parents routinely assume more than exists. A linked parent can set quiet hours (blocking access on a schedule) and study hours (forcing new chats into study mode). A linked parent cannot read the teen’s conversations. Safety notifications exist for certain risk categories. That is the actual scope.

TechCrunch’s coverage on launch day included the caveat that matters most: it is unclear how hard these systems are to escape if a teen decides not to cooperate, and “teens are incredibly adept at working around parental controls.” Treat that as the operating assumption, not a criticism.

Age prediction AI signals: how behavior becomes an estimate

Here is the mechanism, stated plainly. The system is a classifier: it takes features about an account and outputs a probability that the account belongs to someone under 18. It is trained on accounts whose ages are known, learns which feature patterns correlate with each group, and applies those patterns to accounts whose ages are not known.

Think about why each documented signal carries information.

Account age. An account created three weeks ago belongs to a different population than one created in 2023. Adoption waves differ by age group, and a brand-new account has no history to contradict a guess.

Time-of-day activity. This is the strongest intuitive signal. A person who is consistently active at 3 p.m. on weekdays and quiet at 10 a.m. is living on a school schedule. Adults with office jobs show the inverse. Someone active at 1 a.m. on a Tuesday is either a night-shift adult or a teenager who should be asleep, which is precisely why no single signal decides anything.

Longitudinal usage patterns. How does the account’s behavior change over weeks? Vocabulary complexity, session length, request types, whether usage collapses in June and resumes in September. That last one is the school calendar, and it is a fingerprint.

Stated age and conversation topics. If someone mentions their grade, their homework, or a teacher, that is direct evidence. Reports indicate the topics a user raises factor into the estimate.

Now the design decision that parents should actually weigh. This classifier faces a choice about which error to make. Calling an adult a minor is annoying: they lose access to some content and must verify with a selfie through Persona. Calling a minor an adult means a 14-year-old gets adult-mode responses. OpenAI stated it “defaults to a safer experience when confidence is low.” In classifier terms, they tuned the threshold to accept more false positives on adults in exchange for fewer false negatives on kids.

That asymmetry is the whole story, and it is a genuinely reasonable engineering choice. It also has predictable consequences. Adults whose behavior resembles a teenager’s, including night-shift workers, people with new accounts, and people who mostly ask homework-shaped questions, will get misclassified. And a teenager who wants adult mode has a documented target: use the account in adult-shaped ways, or obtain an adult’s verified account. The system raises effort, it does not build a wall.

For context on how hard this problem is in general: Australia’s Age Assurance Technology Trial, whose final report was published on August 31, 2025, tested 48 providers across age verification, age estimation, and age inference. Its conclusion was that age assurance can be done privately and effectively with no substantial technological barrier, and that there is no one-size-fits-all solution. That is the state of the art, from the most thorough government trial available: workable, imperfect, method-dependent.

How to Teach Your Kid About Age Prediction

Ages 5–8: The guessing game about you

Ask your kid to guess your age from three clues you give: what time you wake up, what you had for breakfast, what you did on your phone this morning. Then let them guess a grandparent’s. They will do surprisingly well, and they will notice which clues helped. Say: computers do this with the clues that people leave when they use an app. That is a complete and accurate explanation at this age.

Ages 9–12: Build the signal list

Ask your kid to list every clue a website could use to guess their age without ever asking. Most kids will produce five or six: when I use it, what words I use, what I ask about, how long I have had the account, what apps are on the phone. Then show them OpenAI’s actual documented list and compare. Kids find it satisfying to have guessed correctly, and it converts a vague anxiety into a concrete, bounded mechanism.

Ages 13+: Argue both sides of the threshold

Give your teen the real dilemma: the classifier must choose which mistake to make more often. Have them write a short argument for tuning it toward “assume teen” and another for “assume adult,” then state which they would ship and why. This is applied ethics with a concrete decision, and it is far more productive than a lecture about honesty. Follow with the practical question: what would you do if it guessed wrong about you, and what would you refuse to do?

The question to ask: “If a computer had to guess your age from how you use your phone, which single clue would give you away first?”

The signal table: what it sees, what it means, where it fails

Signal	What it is	Why it correlates with age	How it fails
Account age	How long since the account was created	New accounts skew younger in adoption waves	A new adult account looks young
Time-of-day activity	Hours and days of typical use	School schedules are distinctive	Night-shift adults, insomniac adults
Longitudinal patterns	How use changes across weeks and months	Summer breaks, term rhythms leave marks	New users, irregular users
Stated age	What the user typed	Direct evidence when true	Easily and commonly misstated
Conversation topics	Subject matter raised	Homework and grade mentions are strong tells	An adult teacher or tutor looks like a student
Verification (Persona)	Selfie-based identity check	Establishes adulthood for restored access	Requires handing biometric data to a vendor

The bottom row deserves its own thought. The remedy for a wrong guess is giving a third-party company a photo of your face. For an adult, that may be an acceptable trade. For a family thinking carefully about biometric data, it is a real cost, and it is fair to weigh it before you need it.

What to actually do at home

Decide your family’s stance on selfie verification before you need it

If your teen’s account gets misclassified, or if you as a parent get pushed into verification, you will be deciding under pressure. Decide now: are you willing to submit a selfie to a third-party identity vendor to unlock a chatbot? There is no wrong answer, and there is a wrong time to decide.

Assume the placement is correct and useful, not a punishment

If a kid lands in ChatGPT for Teens, the defaults are better for learning: Study Mode instead of finished answers, no romantic language, no simulated feelings. Common Sense Media’s August 18, 2026 survey of 1,017 teens found 70 percent use AI for schoolwork, and of those, 63 percent get an answer. Study-mode defaults are aimed squarely at that number. Frame the teen mode as the version that makes the tool useful rather than the version that limits it.

Set quiet hours and study hours, then say out loud what they do not do

Quiet hours block access on a schedule. Study hours force study mode. Neither shows you the conversations. Telling your teen precisely that builds more trust than vagueness, and it is also simply accurate. Our guides on ChatGPT parental controls and how ChatGPT guesses under 18 go step by step.

Expect the personal-device workaround

The same Common Sense survey found 44 percent of teens had a tool blocked at school and 59 percent of those switched to a personal device. Age prediction is an account-level system; a new account on a personal phone is a new account with no history. The realistic goal is raising friction and keeping the conversation open, not sealing every gap.

Teach the meta-lesson, because it generalizes

Every platform your kid uses is estimating things about them from behavior: age, interests, mood, purchasing intent. Age prediction is unusual only in that the estimate is documented and its purpose is stated. A kid who understands this one example has a template for understanding recommendation feeds and ad targeting.

What not to do

Do not coach your kid to game the classifier. Beyond the obvious, it removes protections that exist because of real harm cases, including the lawsuits that preceded this product. And it teaches that the response to a safety system is to defeat it, which is a habit with much worse applications later.

What to Watch For Over the Next 3 Months

• Week 4: Your teen can name two signals the system uses and explain why it errs toward “teen.”
• Month 2 red flags: A new account appears on a personal device; or your teen’s ChatGPT suddenly stops showing teen-mode behavior without your involvement.
• Month 3 self-check: Have quiet hours and study hours actually held? If your teen routed around them, that is information about the rule, not just about the kid.

Frequently Asked Questions

Does ChatGPT scan my kid’s face to estimate age?

Not for the prediction itself. The prediction uses behavioral and account signals. A selfie check through the vendor Persona appears only as the remedy when someone contests being classified as under 18.

How accurate is it?

OpenAI has not published an accuracy figure, which is worth noting. The company states the system defaults to the safer experience when confidence is low, which implies meaningful uncertainty. Australia’s 2025 Age Assurance Technology Trial found across 48 providers that age assurance works but with no one-size-fits-all solution.

Can my teen just lie about their age?

They can state any age, and stated age is only one signal among several. Behavior can contradict a claim over time. It is a friction system, not an identity system. Treat it accordingly.

Do I get to read my teen’s chats?

No. OpenAI’s parental controls explicitly do not provide conversation access. You get scheduling controls and certain safety notifications. If you want to know what your teen talks to the model about, the mechanism is asking them.

What happens when my teen turns 18?

The stated age and the behavioral signals both shift over time, and accounts can be verified. Expect the transition to be handled through verification rather than automatically on a birthday.

Is this only ChatGPT?

No. Age assurance is spreading fast, driven by California’s SB 243 and SB 1119, a dozen-plus state laws on companion chatbots, and the FTC’s 6(b) inquiry into seven AI chatbot companies regarding minors. Expect prediction, estimation, or verification on most platforms your kid uses within a couple of years.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. OpenAI. “Our approach to age prediction.” https://openai.com/index/our-approach-to-age-prediction/
2. OpenAI. “Introducing ChatGPT for Teens: Built for learning, backed by protections.” (2026, August 18). https://openai.com/index/chatgpt-for-teens/
3. OpenAI Help Center. “Age prediction in ChatGPT.” https://help.openai.com/en/articles/12652064-age-prediction-in-chatgpt
4. TechCrunch. (2026, August 18). “OpenAI launches a safer ChatGPT for teens.” https://techcrunch.com/2026/08/18/openai-launches-a-safer-chatgpt-for-teens-years-after-teens-started-using-it/
5. Common Sense Media. (2026, August 18). “Teens in the AI Era: Schoolwork and the Skills That Matter.” https://www.commonsensemedia.org/research/teens-in-the-ai-era-schoolwork-and-skills-that-matter
6. Australian Government, Department of Infrastructure, Transport, Regional Development, Communications, Sport and the Arts. (2025, August 31). “Age Assurance Technology Trial: Final Report.” https://www.infrastructure.gov.au/department/media/publications/age-assurance-technology-trial-final-report
7. California Legislature. “SB 1119: companion chatbots, children’s safety.” https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260SB1119