UK to scan asylum seekers' faces for age checks despite knowing the technology is flawed

The UK Home Office will roll out an AI-driven facial-age-estimation system for asylum seekers by the end of the year. Internal pilot reports unearthed by Ars Technica through Freedom of Information requests show the technology does not meet basic accuracy standards.

The system uses a neural network to estimate age from facial features — face geometry, skin texture, bone proportions. Among asylum applicants without a valid identity document such as a birth certificate or passport, the existing process requires a social-services assessment — a detailed biosocial evaluation by qualified social workers that takes on average six weeks. The new technology claims to deliver an outcome in seconds.

Accuracy is the issue. Pilots of three independent vendor systems contracted by the Home Office produced misclassification rates of 22 to 31 per cent on a sample of 16- to 19-year-olds. More worryingly, the errors are not symmetric: the systems are more likely to label minors as adults than the reverse, which is the direction that strips child-protection rights from those affected.

For minor asylum applicants the consequences are large. A child-classified applicant is housed in safeguarded children's accommodation, accesses education services and cannot be detained in adult facilities. Being misclassified as an adult risks children being sent to adult detention centres — a category of placement that is itself prohibited under UK law.

Cathryn Costello, a refugee-rights academic at the University of Liverpool, told Ars Technica: "The Home Office has contracted three vendors and has not published the internal accuracy thresholds of any of them. That tells you the foundational scientific legitimacy of the system is missing. The socio-biographical history of a person whose age is otherwise unknown carries more information than the algorithm does."

The European Court of Human Rights, in its 2019 ruling against a Dutch facial-age-estimation system, held that using such systems without sufficient communication with the applicant violated personality rights. The UK left the EU in 2020, so that ruling is not directly binding, but the same principles are preserved in the Human Rights Act.

The Home Office's case for urgency: 25,000 to 30,000 "age dispute" cases a year, at an average social-services assessment cost of £13,000 per case, yield a total annual burden of £350 to £400 million. The AI system reduces per-case cost to roughly £35. The savings case is real; what scientists are arguing is that those savings are being bought at the cost of child-protection rights.

Germany withdrew a similar system in 2024 after pilot trials. The Federal Office for Migration and Refugees found the system was three times more likely to misclassify applicants of sub-Saharan African origin than those of European origin. Germany has not removed AI-based age estimation from trials altogether, but has reverted to biomedical methods such as dental X-rays and wrist MRI — which are themselves imperfect, with an 8 to 12 per cent error rate.

The Refugee Council, a UK child-protection charity, said in a statement: "If the figures in the Ars Technica report are correct, the system risks sending thousands of children a year into adult detention. That is non-compliance with the UK's obligations under the 1989 UN Convention on the Rights of the Child." The organisation said it is preparing legal challenges.

The broader technology-policy question is wider. Is it defensible to deploy an AI system of known limited accuracy in decisions whose consequences are not recoverable — given that time a child spends in adult detention cannot be undone? The UK Centre for Data Ethics and Innovation, in its 2025 report, advised that in high-stakes settings AI should be positioned as a decision-support tool rather than a decision-maker — a position the Home Office's announcement does not yet clearly accommodate.

The Home Office says the system will operate as "first-stage triage, not final decision" — but the announcement is unclear on the triggers for human review, the AI confidence threshold considered safe, and the route to appeal an AI assessment. Calls for parliamentary scrutiny before the year-end launch are growing.