What AI age prediction is — why ChatGPT kept an account adult after it said 'I'm 13'
AI age prediction estimates how likely an account is to belong to someone under 18 from usage signals rather than an ID document. OpenAI began rolling it out on ChatGPT in January 2026, weighing how long an account has existed, what times of day it is active, usage patterns over time and the age entered at sign-up. Accounts judged likely to be minors — or impossible to judge — get teen protections, and wrongly flagged adults can restore full access with a live selfie or government ID checked by verification firm Persona. Because it works on probabilities, it makes mistakes: in Common Sense Media's October 2026 test, an adult account that repeatedly said it was 13 was never moved to teen mode over a week
The three lines
- Definition — estimates minor status from active hours, account age, usage patterns and stated age, without ID
- How it works — likely minor or unclear → protections on; misflagged adults clear it with selfie or ID (Persona)
- Limits — it does not react to one statement; 'I'm 13' repeated for a week triggered no switch
Key questions
- What is AI age prediction
- **A model that calculates the probability a user is under 18 from behaviour instead of ID.** | Item | Detail | |---|---| | Inputs | Account age, active hours, long-term usage patterns, stated age | | Output | Likely minor / likely adult / unclear | | When unclear | Stricter settings applied (OpenAI) | | Fixing errors | Persona selfie or ID check | | On ChatGPT since | January 2026 |
- How does ChatGPT verify age
- **Two steps: predict from behaviour, then verify with selfie or ID if challenged.** | Step | Method | Who | |---|---|---| | 1. Predict | Usage-based age prediction model | OpenAI | | 2. Verify | Live selfie or government ID | Persona (third party) | | Result | Protections removed if 18+ confirmed | — | | Data retention | Deleted by Persona within 7 days (per OpenAI) | — |
- How accurate is AI age verification
- **It depends on the method, and errors cluster near the age-18 line.** | Method | Strength | Weakness | |---|---|---| | Self-declaration | Simple | Defeated by lying | | Behaviour-based prediction (ChatGPT) | No ID needed | Slow to build signal; misclassification | | Facial age estimation | Result in seconds | Errors near threshold; gaps between groups | | ID check | Most accurate | Privacy burden; users drop off |
ChatGPT heard "I'm 13" over and over and still left the account set to adult. That was a finding in Common Sense Media's October 2026 test ("ChatGPT for Teens rated 'Unacceptable Risk'"). It looks like a bug, but it shows how AI age prediction is designed. ChatGPT judges age from long-run usage patterns, not from a single remark. If it flipped settings on one statement, a single line of text could flip them back. This explainer covers what age prediction looks at, where it fails and how it compares with other ways of checking age.
1. What it looks at
OpenAI began rolling out age prediction on ChatGPT in January 2026. Rather than trusting the birth date entered at sign-up, it also looks at how the account is actually used.
| Signal | What it measures | Why |
|---|---|---|
| Account age | How long since the account was created | New accounts carry little information |
| Active hours | When in the day the user is usually on | School-day rhythms |
| Long-term patterns | Topics and style over weeks | More stable than one remark |
| Stated age | Birth date given at sign-up | A starting point, but can be false |
The model combines these into "likely minor," "likely adult" or "unclear." OpenAI says that when it cannot tell, it applies the stricter setting — teen protections. Those block content such as graphic violence, sexual roleplay, depictions of self-harm and extreme beauty standards.
2. When it's wrong — selfies and IDs
| Situation | Effect | How it's corrected |
|---|---|---|
| Adult flagged as minor | Protections on; some answers limited | Live selfie or ID via Persona |
| Minor treated as adult | Full adult experience, no protections | Stays until parent linking or more signals accumulate |
| Unclear | Protections on | Adults verify with selfie or ID |
The two errors are not equal. An adult flagged as a teen is inconvenienced and can fix it with a selfie. A teen left as an adult gets no protection at all — the error Common Sense Media highlighted. In the same test, the share of crisis prompts answered with a hotline number fell from 33% before the teen launch to 23% after.
This explains why OpenAI built the system to react slowly. If "I'm 13" switched teen mode on instantly, it would be hard to refuse "I'm an adult" switching it off. So the company trusts long-term patterns more, and pays in speed: the test sent about 1,000 prompts over a week without a single switch. OpenAI says the test may have run before features were fully active.
3. How it compares with other methods
There are four broad ways to check age. The more accurate, the heavier the privacy burden; the easier, the easier to fool.
| Method | Accuracy | Privacy burden | Ease of cheating | Typical use |
|---|---|---|---|---|
| Self-declaration | Low | Minimal | Easy | Birth-date fields in most apps |
| Behaviour-based prediction | Medium (rises over time) | Usage analysis | Moderate | ChatGPT, YouTube |
| Facial age estimation | Medium–high | Face image | Hard | Selfie age-check services |
| ID verification | High | Large | Very hard | Persona and similar vendors |
The US National Institute of Standards and Technology (NIST) regularly evaluates facial age estimation. It measures how many years estimates are off on average, and separately how often real 14–17-year-olds are wrongly passed as 18 or older. The threshold is the hardest part: estimating a 17-year-old as 20 or as 14 is the same three-year error, with opposite consequences. That is why deployments often use a buffer, such as sending anyone who looks under 25 to a stronger check even when the legal line is 18.
| NIST metric | Meaning |
|---|---|
| Mean absolute error (MAE) | Average gap between estimated and real age |
| Acc(3) | Share of estimates within ±3 years |
| False positive (minors passed) | Real minors judged above the threshold |
| False negative (adults blocked) | Adults judged as minors |
Behaviour-based prediction has no comparable public benchmark yet, and OpenAI has not published accuracy figures for ChatGPT's model. When YouTube introduced a similar approach, adults complained of being wrongly flagged as minors.
4. What remains and what is unconfirmed
- Accuracy disclosure: OpenAI has not published misclassification rates, so there is no external yardstick.
- Retest: Common Sense Media and OpenAI dispute the method; no retest with fully activated accounts has been published.
- Data deletion: Persona deleting selfies and IDs within seven days is OpenAI's description.
- Other countries: Whether the same signals and process apply everywhere was not confirmed.
- Related: Who is responsible when AI acts on its own is covered in "What AI agent legal liability means."
Sources
- BleepingComputer — OpenAI rolls out age prediction model on ChatGPT to detect your age
- Decrypt — OpenAI deploys age prediction to restrict teen access on ChatGPT
- The Decoder — OpenAI rolls out age prediction to apply teen safeguards in ChatGPT
- Android Authority — ChatGPT age prediction
- NIST — Face Analysis Technology Evaluation (FATE) Age Estimation and Verification
- Biometric Update — Beyond accuracy: what NIST's latest age estimation results mean for age assurance
- Common Sense Media — ChatGPT claims to have a safety net for teens. We watched it fail.