Updates

The age check does not end at the selfie: the system keeps watching

Roblox has processed 338 million age-check records — and built a layer that keeps assessing whether an account's behaviour still matches its estimated age. The reason is blunt: accounts lent to a younger sibling, and accounts probed in bad faith.

by Gustavo Cantino5 min readBeginner

Produced with AI from the sources cited below and checked by automated verification against those same sources. After it was published, the editor read the full text. · Reviewed by Gustavo Cantino, 11/11/26 Editorial Policy

Typographic plate with the Updates category in large type, the beginner level below it, and a field of vertical cyan bars on a dark background.

When you do the Roblox age check, it feels like the story ends there: you took the photo, you got the band, done. That is not what happens. The platform published a technical piece explaining that there is a second layer, which keeps assessing over time whether an account's behaviour still matches its estimated age.

This piece explains why that exists, what it means for your account, and where the system acknowledges its own limits.

The short answer

The first layer is the one everybody knows: facial estimation, or an ID. The system has already processed 338 million age-check records. The technology was chosen because it is tested and certified by third-party laboratories and proved more accurate than self-declared age — which was the previous method.

But the company describes that check as necessary and insufficient, for two concrete reasons:

  • Accounts can be shared or handed to a younger sibling. The person who verified may not be the one playing six months later.
  • Somebody may probe the defences deliberately. Verifying once and then behaving differently would be a predictable route.

So a machine learning layer was built that assesses whether an account's behaviour still matches its estimated age. It is not a second photo: it is a reading of patterns over time.

What it changes for your account

In practice, for most people, nothing changes — and that is the intention. But three things start to make sense:

  1. Lending your account has consequences. If you verified and handed the login to a younger sibling, the account's behaviour stops matching the registered age. That is literally the case the system was built to catch.
  2. An account is individual. Sharing a login was always a bad idea for security; now it is also an age-classification problem.
  3. A band is not permanent by decree. It came from a check and is monitored afterwards. That means it can be revised.

The practical recommendation that follows is dull and simple: one account per person. If somebody younger is going to play, let them play on their own account — which will have the right band, the right protections and the right parental controls.

The two layers of the age check

The table separates the layers, the cases that motivated the second, and the limits the company acknowledges.

The third part is the most honest thing in the technical piece and deserves reading. The company states that age boundaries are not behavioural cliffs: parents may enjoy the same games as their children, and language varies by culture, context and community as well as by age. And it concludes that no single signal captures that nuance — which is why the system combines several.

That is the opposite of a promise of infallibility, and it is good that it is.

The catches

Behaviour is not age. An adult who plays the same things as their child is not a problem, and the text acknowledges it. Any system of this kind lives with false positives.

A lent account is the central case. If you do this, know that it is exactly what the system looks for.

The assessment is ongoing, not a one-off. There is no announced re-verification moment. It is monitoring over time.

It does not replace the initial check. The two layers coexist: one establishes the band, the other checks whether it still makes sense.

No behavioural data is published per account. The text describes the system, not what it saw in you. There is no panel where you look up your own history.

338 million is a cumulative volume. It is the total of records processed, not the number of people verified — the same account can generate more than one record.

What to do next

There is one concrete action: do not lend your account. If a younger sibling wants to play, create their account. It takes five minutes and settles the age band, the protections and the guardian's controls in one go.

And if you are a guardian, this piece is a useful argument in a conversation that usually goes nowhere. "Do not lend your login" sounds like a rule invented by an adult; "the platform watches for exactly this and your sibling will end up in the wrong age band" is a reason. The second works better because it is true and checkable.

One last reading of what this system represents, beyond the mechanics. A one-off check is a photograph; what the platform built alongside it is closer to a running estimate that can be revised. That is a more honest design for a problem that has no clean answer — people share devices, siblings borrow logins, and a face at one moment says nothing about who is typing six months later.

It also explains why the company spends a paragraph of a technical piece listing what the system cannot do. A model that assumed age determines behaviour would be wrong about every parent playing with their child, and confidently so. Stating the limit up front is what allows the rest of the design — combining many weak signals instead of trusting one strong-looking one — to make sense.

There is a reading of this that matters for anyone who worries about being wrongly judged by it. A system built on the explicit premise that no single signal is reliable is, by construction, a system that does not act on one odd week of behaviour. That is not a guarantee, and the piece does not offer one — but it is the difference between a rule that triggers on a pattern and a rule that triggers on an incident.

And for the ordinary player, the conclusion is smaller than the machinery suggests. You are not expected to behave like your age band, whatever that would mean. You are expected to be the person who verified. The only behaviour that reliably breaks that is handing the login to somebody else — which is also, separately, the fastest way to lose an account.

If there is one sentence to keep from the whole technical piece, it is the one about cliffs. Age is a line drawn through a continuum, and any system that pretends otherwise will be confidently wrong about the people standing near the line. Acknowledging that is what makes the rest of it trustworthy, and it is the part most companies leave out.

Sources

How we correct mistakes

Ready-to-use artifact

The two layers of the age check
| Layer | What it is | When it acts |
|---|---|---|
| Initial check | facial estimation, or an ID | once, at verification |
| Ongoing assessment | a model looking at the account's behaviour | over time |

| Why the second layer exists | Case |
|---|---|
| A shared account | more than one person using the same login |
| An account handed on | the login given to a younger sibling |
| An attempt to game it | somebody probing the defences deliberately |

| Limit the company acknowledges | Detail |
|---|---|
| Age boundaries are not cliffs | parents may enjoy the same games as their children |
| Language varies | by culture, context and community, as well as by age |
| No single signal is enough | which is why the system combines several |

| Scale | Value |
|---|---|
| Age-check records processed | 338 million |

Information checked against the official technical piece of 04/08/2026, verified on 23/09/2026. Source: about.roblox.com/newsroom/2026/08/beyond-selfie-how-roblox-age-assurance-system-helps-keep-age-checks-current-
In this article

Keep reading