In this blog series, our CEO Robin Tombs will be sharing his experience, whilst focusing on major themes, news and issues in the world of identity verification and age assurance.
This month, Robin talks about Yoti’s performance in Australia’s Age Assurance Technology Trial, New York’s SAFE for Kids Act and public perceptions of the Online Safety Act.
Yoti’s facial age estimation technology achieves independent validation and success
Yoti has waited a long time to see independent, high-profile and well-funded technology trials of our facial age estimation technology. We’re delighted to read the results which have just been published in Australia’s Age Assurance Technology Trial (AATT) report.
Yoti released our first facial age estimation model in December 2018 and published the performance results. These results were based on testing against over 100,000 images with ground truth ages across a range of 13-60 year olds, split by year of age, gender and skin tone.
Whilst almost all stakeholders welcomed our transparency, some understandably were sceptical of our internal testing results. They wanted independent validation.
Confidence in our technology improved from Autumn 2023, when the US National Institute of Standards and Technology (NIST) started publishing evaluation results for the Yoti March 2023 (‘Yoti-001’) facial age estimation model. Currently, NIST judges Yoti’s September 2024 (‘Yoti-003’) facial age estimation model as the most accurate for estimating the ages of 13-16 year olds.
But today, the AATT report shows that Yoti’s September 2024 model is the most accurate model for its test set of 12-23 year olds. The model achieved a mean absolute error of 1.80 years for this group. Key findings from the report include:
- Yoti’s Target of Evaluation is at Technology Readiness Level 9 – the highest level.
- Manual functional testing confirmed the system’s interoperability, robustness and privacy aspects.
- Automated lab testing showed high accuracy for age gates of 13 years and 16 years. The true positive rates were found to be consistently above 94% from age 13 upwards and the mean absolute error for ages 13-20 was under 2 years.
- There were very few usability issues arising from the Mystery Shopper trials. While the response time wasn’t as fast as some of the other systems tested, users reported that the system met or exceeded their expectations in terms of task completion time. The majority of users found no issues when completing the evaluation task and reported the system was either ‘easy’ or ‘very easy’ to use.
- 80% of users were either ‘satisfied’ or ‘very satisfied’ with the experience and the majority had no concerns about privacy.
- 66% of users reported accurate age estimation results, which is consistent with the overall average accuracy of 70% for the mystery shopper estimation task. We note that some users expected the system to estimate their age exactly (to within a few months). As a result, they reported “no” for accuracy without considering any margin.
The AATT tested many important functions and Yoti passed every one. The results should reassure stakeholders that Yoti’s facial age estimation does not store personally identifiable information (PII) and is capable of detecting paper, screen and mask attacks. It also works effectively for age gating Indigenous (Aboriginal and Torres Strait) Australians at 13+, 16+ and 18+.
Our June 2025 model shows further improvement. The mean absolute error is now 1.60 years for 12-23 year olds and just 1.04 years for 14-18 year olds. This is based on a test set of around 125,000 images from across the globe.
New York sets high standards for age assurance methods
It’s super interesting for Yoti and all followers of age assurance to note that the Office of the New York State Attorney General has just submitted smart, clear details on commercially reasonable and technically feasible methods of age assurance. These are part of the draft Stop Addictive Feeds Exploitation (SAFE) for Kids Act.
Crucially, the Office of the New York State Attorney General has provided key minimum performance levels:
- Minimum accuracy levels for age assurance methods for users aged 18+
- A minimum 98% detection rate for method circumvention.
They have also clarified the following:
- At least one age check method must meet a ‘total accuracy minimum’.
- Operators must implement data minimisation and deletion practices.
- Encryption and security controls for age assurance data must meet industry standards and be independently tested.
- Every age assurance method must be certified annually by an accredited independent third party under international standards (such as ISP/IEC 27566, IEEE 2089.1 or others)
- Independent testing must cover accuracy, false positives and negatives, inconclusive outcomes and circumvention detection. It should also verify data deletion, data minimisation and data security practices.
- Where an operator doesn’t have actual knowledge that a user isn’t a minor, the operator must use certified age assurance methods to confirm adult age status.
- Where age check methods have variable settings or options, operators must only use settings or options that have been certified.
- Zero-knowledge proof of age assurance can be an effective age assurance method that maximises user privacy and data security.
- Zero-knowledge proof allows users to transmit their proof of age status to operators without sharing personal details.
- Double-blind age assurance methods can transmit proof of age status while ensuring that no-one other than the user is aware of both their personal information and the identity of the operator requesting the proof of age status.
The New York Attorney General is now consulting on these recommendations.
This is, by far, the clearest and smartest set of age-checking regulations we’ve seen, delivering privacy-preserving, highly effective age assurance. I suspect they’ll be copied by many age verification regulators around the world over the next 1-3 years.
Clearly, not all the age checks being performed by businesses in the UK market since 25 July 2025 meet New York’s minimum standard. But Ofcom is yet to publish any scientific minimum performance requirements and is not, as of now, investigating any businesses using age check methods that would not meet New York’s standards.
Despite opposition, most Brits support online age protections
In a Sky News article from a month ago, some teenagers expressed that algorithms are no longer pushing as much porn or harmful content into their feeds. It’s not surprising that about half a million people have signed a parliamentary petition to repeal the Online Safety Act, given the vocal opposition from a minority of adults who strongly oppose the law and its age assurance requirements.
The Science and Technology presenter Mickey Carroll concluded the video by questioning whether the government might eventually need to repeal the law.
It’s not uncommon for many adults to dislike certain UK laws, but that does not usually lead to those laws being scrapped, unless a new party is elected that has campaigned to scrap a specific law.
A clear majority of UK adults (not just parents) across all four of the largest political parties support age assurance for controlling access to porn and illegal content such as self-harm and suicide-related material. Understandably, many are concerned about the need for privacy protection, VPN circumvention and the law’s potential effectiveness.
The article did not mention that Yoti, one of several age verification providers, has performed millions of UK age checks in the last month, with many UK adults consenting to these age checks.
I’m often told by opponents of age assurance that many Brits aren’t intelligent enough to understand the risks of supporting this law. But over the last 7.5 years, when offered the choice, over 80% of Brits have consistently chosen to use Yoti’s facial age estimation to prove they’re over 18 without needing to submit identity documents, mobile numbers or credit card information.
A further 15% are choosing to either have their age estimated in the Yoti Digital ID app once or add their identity document to the app so they can easily and anonymously share their 18 age credential in the future.
Based on this evidence, it seems that Brits are more savvy than some critics give them credit for!
By design, we have no idea which individuals in the UK have used Yoti’s facial age estimation or the Yoti Digital ID app to prove their age. However, it’s very likely that between 10 and 20 million Brits have done so in the past 7.5 years, and we rarely receive complaints (apart from some underage kids not liking that it’s very hard to ‘beat’ these two age checking methods).
Many more millions of Brits will need to regularly prove their age on their phones when the alcohol licensing laws change in December.
Never say never with government changes, but I doubt this current government will repeal the Online Safety Act as it had strong cross-party and electoral support.
