AI Age Estimation Is Weakest at the Legal Thresholds It Is Meant to Enforce

Artificial intelligence systems are increasingly used to estimate users' ages for platforms requiring age verification, but they rely on statistical guesses rather than confirmed identity documents. Independent testing, including Australia's 2025 Age Assurance Technology Trial covering over 60 solutions, found that even the best facial systems carry error margins of two to three years around key legal age thresholds such as 13, 16, and 18. This means a system accurate on average can still misclassify a 16-year-old as 19 or a 20-year-old as 17 near the boundaries that regulators most care about. Beyond accuracy, these errors are not distributed evenly across demographic groups, making age estimation a bias issue as much as a technical one. Behavioural inference methods, such as ChatGPT's approach of reading usage patterns rather than faces, raise similar concerns by inferring a protected characteristic from observable data without any authoritative check.