The unappealable number

A leaked British government report, finished in April 2025, tested seven facial age-estimation algorithms against more than 2.5 million images. The best one was off, for female Sub-Saharan Africans, by an average of 4.6 years — enough to read a 13-and-a-half-year-old girl as an adult. The Home Office bought a system anyway. Starting in 2027, Britain will scan the faces of asylum seekers at its border and let an AI guess whether they are children.

The detail that matters isn't the error rate. It's that the government knew the error rate and proceeded — because the people the system gets wrong are the people who can't do anything about being gotten wrong. Accuracy is a constraint only when the misclassified can push back. Strip that away and "we know it's flawed" stops being a confession. It becomes the design.

Accuracy is a constraint only when the misclassified can push back.

The errors aren't random. They run in one direction and land on one group. The tested system tended to call a 17-year-old an adult. Cognitec, the German firm the UK paid more than $400,000, misclassified twice as many 16-year-olds as over-18 on the low-quality photos taken at borders as it did on clean visa shots. Sixteen-year-olds from West Africa were flagged as adults more often than Eastern Europeans the same age. Sub-Saharan Africans are the largest group the UK assesses. Even trauma counts against you: the report found the stress of the journey appears to age a face upward in the model's eyes.

And the people it ages upward lose the most. A child wrongly classed as an adult is stripped of safeguarding protections and can be placed in adult detention. There is no symmetric harm running the other way — an adult read as a child is the exact failure the program exists to prevent. The Home Office sold the tool as "cutting-edge AI" to "crack down on fake claims" and stop "adults attempting to game the system." Read that as a spec. The tool is procured to push the count one direction, and its errors oblige.

The Home Office has an answer, and it sounds like the right one: the scan is only an "additional" tool, it won't "replace or overrule human judgment," and "in cases of uncertainty, individuals will always be treated as children until a further assessment is conducted." If that held, the bias wouldn't bite. But the safeguard is doing the rhetorical work, not the procurement. The human judgment the AI is there to "support" already classes 40% of assessed arrivals as adults. That is the number a program sold on stopping fake claims is bought to raise, not to check. And the experts who would have flagged the inaccuracy weren't overruled — they were dissolved. The Home Office shut down its own scientific advisory committee while it was shopping for the algorithm.

Tim Cole, the UCL medical statistician quoted there, calls the scans "hideously inaccurate." You don't disband the committee that agrees with you.

The system doesn't have to work to do its job. Its job is to take a border officer's contestable guess and launder it into a number — one a 13-year-old who just crossed the Channel in a boat has no standing to appeal. The accuracy was never the point. The unappealability is.