A margin of error is inherent in facial age estimation – it is a feature not a fatal flaw. It is designed into how the technology is used, exactly as it is offline. Shops apply policies such as “Challenge 25” because staff cannot judge 18 precisely, anyone who looks under 25 (or sometimes an older limit) is asked for ID. Online systems use the same logic, setting a buffer above the legal age, so a user must be estimated well above the threshold to pass on estimation alone.
Independent testing by NIST, the US government’s measurement laboratory, shows the best systems estimating age with an average error of around two years, and under two years for teenagers. This technology is improving year on year. With a sensible buffer, the realistic failure case is that an adult who is estimated to be too young is asked to complete a second check, not that a child is waved through.