Meta is guessing your age, and it won't say how often its AI gets it wrong

What this is about

On November 9, 2026, Meta announced it is rolling out its AI age detection on Instagram worldwide. The system looks for accounts that probably belong to teens between 13 and 17, even if an adult birthday is on file. It automatically moves those accounts into a "Teen Account," where messages from strangers are restricted, the profile is private and usage time is limited. The technology is also coming to Facebook, and later to Threads.

For Germany, this is less new than it sounds. According to Meta's own blog post from May 5, 2026, automatic classification was already expanded to all 27 EU countries and Brazil back then. What's new is the rest of the world.

The more important question is a different one: how often does a system like this get it wrong, and who does it hit? That's especially relevant right now. The German government is planning a social media ban for under-13s, and the EU wants minimum ages with mandatory age verification. Both assume machines can reliably tell how old someone is.

What Meta reveals and what it doesn't

Meta says it is "encouraged" by the results. An older blog post claims the AI-assisted review delivers "higher accuracy in testing" than humans alone. There's no evidence for that. Meta publishes no accuracy rate and no error rate. There are no per-country figures and no independent audit. TechNode, Medianama and Engadget have all pointed this out.

Much also remains unclear about which signals are evaluated. According to Meta, the AI reads text in profiles, posts, comments and captions. It looks for things like birthday wishes or references to school grade. On top of that come usage behavior and reports from other users.

Since May, an additional image analysis has been checking photos and videos for "general visual patterns." Meta stresses this isn't facial recognition. But Meta says the feature is only available "in select countries," and won't say which. Engadget initially reported that Meta also analyzes bone structure. The outlet later corrected itself: Meta only uses such features to find and remove children under 13. In Australia, Meta also uses the age that Apple stores when a child's iPhone is set up.

If you're wrongly flagged as a teen, you have to prove your age. There are two ways: upload an ID, or record a video selfie that British provider Yoti analyzes using facial age estimation. Meta doesn't say what estimate Yoti has to arrive at for the account to count as adult.

The only hard error limits are in a different document. In August 2026, Meta reached a settlement in the US with attorneys general from 51 or 52 states and territories, depending on how you count. According to analyses by Zyphe and civil liberties group EFF, the agreement governs the "under-18 false positive rate": the share of teens wrongly classified as adults.

For Meta's own technology, the caps in the first year are 14 percent of 16- to 17-year-olds and 7 percent of 13- to 15-year-olds. Later they drop to 10 and 5 percent. For third-party methods, the limits are 10 and 3 percent. For the reverse error, adults stuck in Teen Accounts, the agreement sets no limit, according to the EFF. It only requires a "clear and conspicuous" way to appeal. And these rules apply only in the participating US states.

What can be measured about Yoti

There is measurement data for the selfie appeal. The US standards agency NIST continuously tests age estimation algorithms in its FATE program, using millions of government photos, including visa and police photos. The yardstick is mean absolute error: how many years off the estimate is on average.

Yoti attributes the gap between its own figures and NIST's to image quality. Its own numbers come from high-resolution phone selfies, while NIST works with government photos. In addition, NIST's photos of children come from US consulates in Mexico.

That explanation is plausible. But only what NIST measured is actually verified. NIST also found that errors are higher for women with almost every algorithm. Accuracy also depends on country of origin and image quality. For one test subject, estimates in a phone video varied by several years from frame to frame.

We ran the numbers: what a one-year error means at an age threshold

An error of "just" one or two years sounds small. At a fixed age threshold, it's big. For a rough model calculation, we assume the errors are roughly normally distributed, scattering around the true value like a bell curve. The optimistic case uses Yoti's own figures, the cautious case uses NIST's.

Our own model calculation. The "optimistic" figure for 17-year-olds matches Yoti's own claim: 0.6 percent of 13- to 17-year-olds are estimated at over 21.

Our assessment: with a hard threshold, people just below the line get waved through very often. If you protect against that with a safety margin, known as a buffer, the problem shifts to young adults. That's exactly the buffer Germany's Commission for the Protection of Minors in the Media (KJM) requires for Yoti: since October 2025, to access 18+ content, someone has to be estimated at over 21; before that, it was over 23.

Australia's government trial of age assurance confirms the order of magnitude with real measurements. Within a "gray zone" of two to three years around each age threshold, estimates become unreliable. At the 13 threshold, the systems wrongly rejected 22 percent of 13-year-olds, and still 6 percent of 15-year-olds. False positives were more common for people with darker skin and for 16- to 20-year-olds.

Then there are retries. If someone has a 36 percent chance of getting through on a single attempt, the odds rise to about 74 percent mathematically if they try three times. That's how many attempts per day the US settlement allows. Since the same person tends to get similar estimates, the effect is smaller in practice, but it's there.

Australia: 4.7 million blocked accounts, barely any less use

Since December 10, 2025, social networks in Australia have been off-limits to children under 16. By mid-January, the government says platforms had blocked around 4.7 million accounts. The regulator eSafety surveyed 803 children aged 10 to 15. Before the ban, 85.9 percent used at least one of the affected platforms; three months later, 81.5 percent did. The share with their own account fell from about 52 to 42 percent.

Of the children who kept their account, more than half said they had never been asked their age. 37 percent had given an older age. 18 percent had been wrongly estimated as 16 or older by age checks. Back in March, eSafety criticized platforms for letting under-16s repeat facial estimation until they got a "16+." The regulator also wrote that the method has "higher error rates" near the 16 threshold. Investigations are underway into Instagram, Facebook, Snapchat, TikTok and YouTube. A University of Newcastle study in the BMJ found that around 85 percent of under-16s continue to use social media.

What this means for Germany and the EU

On October 14, 2026, the German federal cabinet adopted key points for a social media ban for under-13s, with tiered restrictions up to 18. The ban is meant as a national interim rule until the EU has its own regulation. There's no draft bill yet. The industry association Bitkom considers a fixed age limit of 13 unsuitable.

On September 17, the European Commission presented the "Kids Act." Under it, children under 13 would not be allowed on social networks. 13- and 14-year-olds would get "mini accounts" through their parents, and their own account only from age 15. For proof of age, the Commission plans its own app, set up using a passport or national ID card. According to netzpolitik.org, comparable tools are permitted if they're similarly accurate and privacy-friendly. Under a non-binding recommendation, all EU member states should offer age verification by December 31, 2026. The 2025 DSA guidelines already consider self-declaration insufficient. There, age estimation is only intended for medium-risk cases and is supposed to be independently audited.

According to netzpolitik.org, the German Ethics Council opposes a ban. It warns that "it cannot be ruled out that age verification technologies will be misused as an instrument of censorship." In March, more than 400 IT security and privacy experts called for age checks not to be introduced without further research. The Chaos Computer Club also rejects social media bans on principle. Appeal partner Yoti has drawn criticism too: Spain's data protection authority fined the company €950,000. That concerned its digital ID app, not its age estimation. Yoti has appealed.

Analysis

AI age estimation is a useful tool for exactly one task: sorting people who are clearly above or clearly below a threshold. For the thresholds at stake politically (13, 15, 16), it's fuzzy. And that applies precisely to the age groups that are supposed to be protected.

Our assessment: on its own, it isn't enough to underpin legal minimum ages. It needs a buffer and an ID check as a second step. Which brings you right back to the ID you were trying to avoid.

Meta makes evaluating all this even harder. A company that automatically classifies millions of accounts but discloses not a single error rate is asking for trust without giving any basis for it. It also remains unclear how many adults Meta keeps stuck in Teen Accounts, which signals are evaluated in Germany, and how quickly appeals are decided. Australia also shows that even mass account blocks achieve little as long as platforms don't check consistently and kids can find workarounds.