The official trial of age assurance technologies that could be used to keep under-16s from accessing social media has found that while there are effective options, errors are inevitable and age estimation tools are more likely to make mistakes for users within two years of 16.
The long-awaited report from the trial found that companies may need to provide multiple “fallback options” such as ID checks in those cases where age estimation technology fails.
The $6.5m age assurance technology trial, run by the UK-based Age Check Certification Scheme (ACCS), tested various types of technology that could be used by social media platforms and adult websites to keep out under-16s or under-18s, respectively, when Australia’s under-16s social media ban comes into force in December.
The report, released by the communications minister, Anika Wells, on Sunday evening, runs to 10 volumes, testing 60 technologies from 48 age assurance vendors. It tested out facial age estimation, age verification using IDs, parental controls, parental consent, age inference and other methods including verification on the phone operating system level.
Sign up: AU Breaking News email
ACCS said in the report it was not the role of the trial to recommend a specific type of technology or whether age assurance should be used at all.
It conducted 28,514 facial age estimation tests with 13 vendors across Australia and found most tests took less than 40 seconds to complete.
It was a “fundamental misunderstanding” of age estimating technology that it could be used without a “margin of error”, the report said: “False negatives will then be inevitable and alternative methods will be required to correct them.”
The report found there is a two- to three-year “buffer zone” for facial age estimation where it is likely that errors in correctly estimating a person’s age will increase, for example if a user is 17 and it estimates them to be under 16, or a 14-year-old told they are over 16.
The report said these users would probably need to be referred for age verification as a fallback method.
“The implementation of buffer thresholds also plays a crucial role in enabling successive validation, a layered approach to age assurance where one method (e.g. age estimation) is followed by a second, more definitive method (e.g. document-based verification) when uncertainty is high,” the report said.
For people aged 16 and 17, false rejection rates remained “above acceptable levels” for facial age estimation, the report found, at 8.5% and 2.6% respectively.
Another major issue identified was that there were accuracy issues for older adults, non-Caucasian users and female-presenting people near the age levels tested, as well as underrepresentation of Indigenous people in the training data.
“Skin tone is a known source of bias in age estimation systems,” the report said.
“As light reflects off a subject’s face and into the camera sensor, lighter skin tones tend to return more detail-rich signals because they reflect more light; darker skin absorbs more light, reducing contrast and lower data availability for the algorithm to analyse.”
The report noted that both Meta and Snapchat already use facial age estimation technology.
Other biometric tests, including hand gestures and voice analysis, were said to be “promising” but still not yet developed enough for wide deployment.
Under the law, providers can use age verification via ID checks as one method but it cannot be the only method offered.
The report also found while most age verification providers kept as little data as possible, there was “concerning evidence” that, without guidance, some providers were “over-anticipating the eventual needs of regulators about providing personal information for future investigations” and were building tools to enable regulators or law enforcement to retrace actions of people to verify their age, which the report said could lead to increased risk of privacy breaches.
Age inference – where a platform estimates a user’s age based on behaviour of the account, interests, friends and other data – was found to be useful to flag likely underage access and support interventions or trigger fallback mechanisms such as age estimation, and is more “low friction” for users who are over the age restriction.
Circumvention using AI and faked images were also something vendors said they had tools to combat, such as using live motion tracking in facial age estimation.
Some vendors highlighted their software could be used to detect if users were using virtual private network connections to dodge the age restriction requirements in Australia. The report found that while it was “promising, this approach raises its own challenges related to accuracy, evasion tactics and the implications for user privacy and cross-border service access”.
Wells said there was no “one size fits all” solution to age assurance. “This trial shows there are many effective options and importantly that user privacy can be safeguarded.”
The federal government was “pushing forward with our mission to keep kids safer online through world-leading reforms”.
