Facial age estimation accuracy is good enough for regulators to accept it, but only with a safety margin. Leading systems typically estimate adult ages to within a few years on average, and errors grow for some ages, demographics and poor-quality images. That is why compliant sites use a challenge age: users estimated well above 18 (for example 21 or 25) pass, and anyone closer to the line is asked for another check.
This guide explains how the technology works, what the error figures actually mean, how buffers and bias testing work, and where European regulators accept or restrict it.
General information, not legal advice.
How AI age estimation works
The user takes a live selfie in the browser or app. A machine-learning model, trained on large sets of face images with known ages, outputs an estimated age, and sometimes a confidence range. The system then compares that estimate with the threshold you have set and returns a pass or fail to your site.
Three components decide real-world quality:
- The estimation model, which determines raw accuracy.
- Liveness detection, which checks that a real, present person is in front of the camera rather than a photo, video, mask or deepfake. France's Arcom requires it for facial methods.
- The decision rules, including the challenge age and what happens to users who fail.
Facial age estimation is not facial recognition. It does not need to identify who the person is, and well-designed systems delete the image straight after the check. It still processes biometric-derived data, so you need a lawful basis and a data protection impact assessment. See our guide to age verification and GDPR.
Facial age estimation error rate: what the numbers mean
The headline accuracy figure vendors quote is usually the mean absolute error (MAE): the average gap, in years, between estimated and true age across a test set. An MAE of two years means that, on average, estimates are two years out, in either direction.
MAE is useful, but it hides three things you need to know:
- It is an average. Some users will be estimated much further from their real age than the MAE suggests. What matters for compliance is how often a minor is wrongly passed at your chosen threshold.
- It varies by age band. Many vendors publish separate figures for 13 to 17-year-olds and young adults. Ask for these, not an all-ages average.
- It depends on the test images. High-quality, well-lit images score better than real selfies taken on cheap phones at night.
For independent figures, the US National Institute of Standards and Technology (NIST) runs an ongoing Face Analysis Technology Evaluation for age estimation and verification, with per-algorithm report cards. Its first report, in 2024, found no single algorithm performed best across every demographic group, and that accuracy depended strongly on image quality. Ask any vendor whether its current model has been submitted to NIST and how it performed, and read the results yourself.
Check the right metric: ask vendors for the rate at which under-18s are passed at your challenge age (often called the false positive or "false accept" rate for minors), broken down by age, sex and skin tone. A low MAE on its own does not tell you this.
Age estimation buffers and challenge ages
A challenge age, or buffer, turns a probabilistic estimate into a safe decision. Instead of passing anyone estimated at 18 or over, you pass only those estimated at or above a higher age, and you route the rest to a second method such as photo ID.
Regulators expect this. Ofcom's July 2026 report on the use of age assurance found that all but two of the adult services using facial estimation applied a challenge-age approach, and described the challenge age as key to the overall effectiveness of the process. In Germany, the KJM has approved facial estimation with a buffer of three years above the age limit, so an 18+ service only passes users estimated at 21 or over.
A worked example (hypothetical numbers)
Imagine a site with 10,000 UK visitors a day reaching its age check. With a challenge age of 25, suppose 70% of adult users are estimated at 25 or over and pass instantly. The other 30% are offered photo ID or a credit card check, and some of them leave. Raising the challenge age to 25 from 21 lowers the risk of passing a minor, but it sends more genuine adults, particularly those in their early twenties, to the slower fallback.
That is the core trade-off. A higher buffer is safer and harder to criticise, while a lower buffer converts better. Test with your own audience and keep a record of why you chose your setting.
Bias testing and fairness
Ofcom's criteria for highly effective age assurance include fairness: a method should not produce materially worse outcomes for particular groups. For facial estimation, that means checking performance across sex, skin tone and age bands.
When you assess a vendor, ask for:
- error and false-accept rates by demographic group, from an independent test lab where possible;
- details of the training data and how it was sourced;
- how often the model is retrained and re-tested, and whether you are told when it changes;
- how the system handles glasses, facial hair, make-up, head coverings and low light;
- a fallback route for anyone the camera cannot process, so no group is locked out.
Certification against an age assurance standard, such as the UK's PAS 1296 or the international ISO/IEC 27566-1, gives extra comfort. Our guide to choosing an age verification provider covers certification in more depth.
Where regulators accept or restrict facial estimation
| Jurisdiction | Position on facial age estimation |
|---|---|
| UK (Ofcom) | Listed as capable of being highly effective, with a challenge age expected in practice |
| France (Arcom) | Accepted within the technical standard, with liveness detection; sites must also offer at least one double-anonymity option |
| Germany (KJM) | Approved within assessed systems, with a three-year buffer |
| Italy (AGCOM) | Checks must be run by independent certified providers using double anonymity; some providers use face-based estimation within that model |
| EU (DSA guidelines) | Verification is recommended for pornography; estimation is recommended for lower minimum ages |
The EU position is the one to watch. The DSA Article 28 guidelines point adult platforms towards verification, so if you serve EU users through a platform in scope of the DSA, plan to offer a verification method alongside any estimation flow. For a wider comparison, see age verification methods compared and the UK detail in highly effective age assurance methods.
Frequently asked questions
Can facial age estimation be fooled?
Weak systems can be fooled by photos, videos or masks. That is why liveness detection matters, and why you should ask vendors how they test against presentation attacks and deepfakes. Regulators expect methods to be robust against easy circumvention.
What challenge age should I use?
There is no single legal figure in the UK. Common settings are 21 to 25 for an 18+ service, while Germany's KJM has worked with a three-year buffer. Pick a setting your vendor's minor false-accept data supports, document the reasoning, and review it as the model changes.
Is facial age estimation GDPR compliant?
It can be, if you process only what is needed, delete images straight after the check, use a processor contract and complete a DPIA. Because face images can be biometric data, take care over the lawful basis and the information you give users.
Next steps
Facial estimation works best as the fast lane in a multi-method flow. Compare vendors that offer it, with credible accuracy data and fallbacks, in our compliance and verification services directory, and see the rest of our age verification guides.