Trial of tech that could be used to keep Australian under-16s off social media finds some errors ‘inevitable’
A government‑commissioned trial of age‑assurance tools for Australia’s upcoming under‑16 social‑media ban found that while many technologies work, errors are unavoidable, especially for users close to the age limit. The report recommends layered verification, including ID checks, to address false p…
The Australian government’s $6.5 million trial of age‑assurance technologies has concluded that errors are inevitable when trying to keep under‑16s off social‑media platforms. The extensive report, released by Communications Minister Anika Wells on Sunday, examined 60 solutions from 48 vendors and highlighted the need for layered verification methods to protect both children and platform operators.
What the trial tested
Run by the UK‑based Age Check Certification Scheme (ACCS), the trial evaluated a wide spectrum of tools that social‑media companies and adult‑content sites could deploy once Australia’s under‑16 ban takes effect in December. The technologies fell into several categories:
- Facial age estimation – algorithms that guess a user’s age from a webcam snapshot.
- Document‑based ID verification – scanning passports, driver’s licences or other government‑issued IDs.
- Parental controls and consent mechanisms – allowing parents to grant or restrict access.
- Age inference – analysing account behaviour, interests and connections to infer age.
- Device‑level checks – leveraging operating‑system data to confirm a user’s age.
In total, ACCS performed 28,514 facial‑age tests across 13 vendors, with most assessments completing in under 40 seconds.
Key findings and error margins
The report makes clear that facial‑age estimation cannot operate without a margin of error. Users who are within two years of the 16‑year threshold are especially prone to misclassification. For 16‑year‑olds, the false‑negative rate was 8.5 %, meaning nearly one in twelve could be incorrectly blocked. For 17‑year‑olds, the rate fell to 2.6 % but remained above what regulators consider acceptable.
These “buffer zones” create a two‑ to three‑year range where the technology’s confidence drops sharply. In practice, the trial recommends that platforms refer users in this zone to a secondary verification step, such as an ID check, before granting or denying access.
Bias and representation concerns
Beyond raw accuracy, the trial uncovered systematic biases. Facial‑age models performed less reliably on older adults, female‑presenting users, non‑Caucasian faces and Indigenous Australians. The underlying training data lacked sufficient diversity, leading to poorer signal quality for darker skin tones—a well‑documented issue in computer‑vision research.
Meta and Snapchat were cited as current users of facial‑age estimation, though Snapchat later disputed the claim, stating it does not employ such technology on its platform.
Alternative and complementary methods
Age inference emerged as a useful, low‑friction tool for flagging likely under‑age accounts. By analysing posting patterns, friend networks and content preferences, platforms can trigger a fallback verification without burdening users who are clearly over the age limit.
Other biometric approaches, such as hand‑gesture recognition and voice analysis, were described as “promising” but not yet mature enough for large‑scale rollout.
Vendors also demonstrated anti‑circumvention measures, including live motion tracking to detect static or AI‑generated images, and detection of VPN usage that might be employed to bypass Australian restrictions.
Privacy and data‑retention issues
While most age‑verification providers limited data collection, the report flagged “concerning evidence” that some were building capabilities to retain extensive personal information for potential future regulator or law‑enforcement queries. This raises privacy risks, especially if data is repurposed beyond the immediate verification need.
Minister Wells stressed that no single solution will satisfy every scenario. “There is no one‑size‑fits‑all answer,” she said, emphasizing the importance of safeguarding user privacy while pursuing the broader goal of keeping children safer online.
What comes next?
With the December deadline approaching, platforms are expected to adopt a layered approach: initial low‑friction checks (like age inference), followed by higher‑certainty methods (facial estimation), and finally definitive ID verification when confidence is low. The government has indicated it will continue to monitor the rollout and may issue further guidance on acceptable error thresholds and privacy safeguards.
Industry observers note that the trial’s findings mirror global challenges in age‑gating digital services. As other jurisdictions contemplate similar bans, Australia’s experience could shape international standards for age‑assurance technology.
Why it matters
The findings determine how effectively Australia can enforce its under‑16 social‑media ban while balancing child safety with privacy and platform usability.
Key points
- Facial age estimation shows high error rates for users within two years of the 16‑year cutoff
- Biases affect accuracy for non‑Caucasian, female‑presenting and Indigenous users
- Layered verification—combining inference, estimation and ID checks—is recommended
- Privacy concerns arise from excessive data retention by some verification providers
Frequently asked questions
What is the Australian under‑16 social‑media ban?
Australia will prohibit anyone under 16 from creating or accessing accounts on most social‑media platforms from December 2025, aiming to reduce exposure to harmful content.
Why can’t a single age‑verification method be used?
No technology is flawless; each method has error margins or privacy trade‑offs, so a combination ensures higher accuracy and protects user data.
Which technologies performed best in the trial?
Age inference and document‑based ID checks showed the lowest false‑negative rates, while facial estimation was useful but required fallback mechanisms.
How does bias affect age‑estimation tools?
Training data lacking diversity leads to poorer performance on darker skin tones and certain demographic groups, increasing misclassification risk.
