Everything you need to know about our facial age estimation technology

profile picture Yoti 8 min read
Woman using Facial Age Estimation to correctly estimate her age

We’re really proud of our facial age estimation technology, but it’s important to us that we tackle some common misconceptions. So we’ve answered some of the most important questions we get asked about facial age estimation.

Woman using Facial Age Estimation to correctly estimate her age

What is facial age estimation?

Our facial age estimation technology accurately estimates a person’s age based on a selfie. We built it to give everyone a secure way to prove their age without sharing their name or ID document.

This privacy-friendly approach doesn’t require any personal details or ID documents, and all images are instantly deleted once someone receives their estimated age – nothing is ever viewed by a human. It can’t link a name to a face or identify anyone. This is the difference between facial analysis and facial recognition.


How does it estimate age?

The technology has been trained to estimate age by looking at facial features in an image. To the technology, the image is simply a pattern of pixels, and the pixels are numbers. Our facial age estimation technology has been trained to spot patterns in numbers, so it learns ‘this pattern is what 16 year olds usually look like’.


Is it facial recognition?

No. The technology uses facial analysis to estimate a person’s age without identifying or recognising any individual.

When estimating age, it doesn’t cross-check people against a big database of faces, it simply estimates the age of the image presented to it. The technology never knows or learns the name or identity of a person.

This has been acknowledged by multiple regulators and has prompted the ICO to update their definition of biometrics and agree that “Yoti’s age estimation tool will not result in the processing of special category data”.


Where did the data come from to train facial age estimation?

Most of it comes from Yoti app users. We use the photo from an ID document and the date and year of birth. We ask them to allow us to use the data to train our technology during the account creation process. Those users can withdraw their approval to use their data for training purposes at any time through the Yoti app.

All our data for under 13 year olds has been collected specifically for the purposes of training the age estimation model and the parents or guardians have freely given consent during a specific data collection exercise.

We have ensured we have obtained a balance of gender identities, images and different skin tones to minimise bias in the algorithm – we have not used any images scraped from the internet.


Is the technology more accurate than humans at estimating age?

It can be difficult for humans to be sure whether someone is over 18 just by looking at them. When guessing the age of another person, we tend to underestimate the age of older people, and overestimate the age of younger people. Our ability to estimate accurately tends to decrease as we ourselves get older.

That’s why policies such as ‘Challenge 25’ exist, which ask customers to prove their age if they appear to be under 25. This builds in a safety buffer of 7 years given the legal age for buying alcohol in the UK is 18.

Facial age estimation can be configured to work with legal age thresholds in a similar way. A retailer using age estimation at self-checkouts can build in a buffer of 7 years, meaning anyone estimated to be under the age of 25 will not pass and will need to prove their age another way.

If we look at the accuracy of the technology when estimating the age of an 18 year old, there is a mean absolute error (MAE) of 1.22. This means an 18 year old could be estimated to be 17 or 19 years – too high or too low. This is why businesses using age estimation can use a threshold, like Challenge 25, to have greater confidence that someone under the age of 18 would not be able to buy an age restricted item.

Another way to measure the accuracy of age estimation is using a True Positive Rate (TPR). This is the probability that an actual positive will test positive, meaning an 18 year old is correctly estimated to be under 23. The TPR for 13-17 year olds correctly estimated as under 23 is 99.65%. This gives regulators a very high level of confidence that nobody underage will be able to access age-restricted content, and the technology can be used in a variety of settings to strengthen age checks – retailers selling age restricted items, adult websites or content providers, and gambling terminals.

We continually measure the accuracy of our technology and improve it.


Is it biased against skin tones?

At Yoti, we take our ethical responsibilities as a company developing new technology very seriously. The data (face image and month and year of birth only) used to train the algorithm is obtained by Yoti in accordance with the UK GDPR during the onboarding process for the Yoti app or using consented data collection exercises.

We train our AI with images from a wide demographic of society and then test how well it estimates people’s age for different genders and skin tones. This allows us to see where it needs to improve so we can train the model in those areas. We also invest heavily to minimise bias and make sure it works well for everyone.

You can see in our white paper that there’s minimal bias across gender and skin tone for 6-12 year olds, with more bias for older adults. While this is something we’re improving, it’s more important that the model is accurate for 18-25 year olds.


Is facial age estimation secure?

We created facial age estimation to give everybody a secure, private way of proving their age. Security, therefore, isn’t just a priority, it’s fundamental to everything we do.

Privacy is a key consideration of facial age estimation – there’s no login required or the need for an account with Yoti. People simply present their face to a webcam or camera on a device, and as soon as their age has been estimated from that image, the captured image is deleted. We have been independently audited both by the ICO and KPMG on this process of deleting images. The ICO’s report is public and KPMG’s is available on request.

We also commission regular external audits of our business and have been certified to meet some of the world’s most stringent security standards, such as ISO27001 and SOC2 Type II.


Is my privacy protected?

Yoti’s facial age estimation is built in accordance with the Privacy by Design principle in the UK GDPR. No individual can be identified by the model and it is designed to minimise the data shared, so instead of requesting an ID document, it just needs an image. This image is never seen by a human and is deleted as soon as the age has been estimated.

We also designed it so that no individual can be identified by the model. For this reason, the ICO has stated that it can be distinguished from facial recognition technology, as it’s not being used for the purpose of uniquely identifying individuals. Instead, it’s used to categorise people by age. They concluded that our age estimation tool will not result in the processing of special category data.


How is facial age estimation currently being used?

Instagram is using the technology to verify the age of users that change their date of birth from under 18 to 18 or over to make sure both teens and adults are in an age-appropriate experience for them. This was first introduced to Instagram users in the US, and has since rolled out to those in the UK, Europe, Brazil and India.

Following the success on Instagram, Meta also introduced the same technology to Facebook Dating to verify that only adults are using the service and to help prevent teens from accessing it. This is initially for people based in the US to ensure only adults can access the 18+ experience.

Social network Yubo uses facial age estimation to check the age of users, ensuring 13-14 years olds are not engaging with adults.

UK supermarkets have trialled age estimation at self-checkouts to give shoppers an easy and quick way to prove their age, without needing to show an ID document or wait for assistance. During the trial, shoppers purchasing alcohol could simply look at a camera on the self-checkout to verify their age before completing their purchase. If the system detected a customer looked younger than 25, they could prove their age through the free Yoti app instead. For those who did not wish to use this technology, they still had the option of showing their ID document to a member of supermarket staff.

Regal Gaming technologies, a leading provider of gaming machines to the leisure and hospitality industry, is testing the technology to create safer and more age appropriate experiences and bring a halt to under 18 players.


Any other questions?

We hope that this answers some of your questions about facial age estimation. If you still have any questions or would like to learn more about our technology, feel free to get in touch.