Internal tests found the best algorithm misclassified some teenage girls as adults by more than four years. The Home Office signed a $400,000 contract for the technology anyway, while 62 civil society groups demanded it be scrapped.
The Home Office had a scientific committee that could have said no. Then it didn't.
That is the structural story behind the United Kingdom's plan to deploy facial age estimation, an AI that guesses someone's age from a photograph, at the asylum border beginning in 2027. The story is not only that the technology is flawed. It is that the body with standing to formally object to a flawed deployment no longer exists, and that the procurement contract was signed before any meaningful external review could land.
Internal testing obtained by a joint investigation from WIRED (reporting by Natasha Ol主持, Maya Share and others), Lighthouse Reports (reporting by Jessica Nicole Costello and others), and The Independent, and republished by Ars Technica (reporting by Dylan Baker and others), makes the technical case concrete. In April 2025, the Home Office ran seven facial age estimation algorithms against roughly 2.5 million images before buying any of them. The best-performing algorithm in that test still showed what the report called "substantial deviations" on Sub-Saharan African faces and tended to predict that a 17-year-old was over 18. For female Sub-Saharan African subjects, the mean absolute error was 4.6 years in that single test, meaning a 13.5-year-old girl could be classified as 18 on a routine scan. The figure is illustrative of the worst-performing cohort on the best-tested algorithm in that one trial, not a generic industry benchmark.[^1]
[^1]: The 4.6-year mean absolute error figure is drawn from a leaked Home Office internal report covering the 2025 trial; the specific image count of "roughly 2.5 million" appears in that same report. The draft's characterization of these figures follows the Ars Technica reporting, which attributes them to the joint investigation's access to the leaked document.
The photo problem is not separable from the demographic problem. The same internal report found that the initial photographs taken at encounter were "routinely worse" than follow-up photographs, and that the study could not disentangle whether the issue was image quality, lighting, or the physical condition of the people being photographed. Sub-Saharan Africans are the largest group of small-boat arrivals to the UK and had the most age assessments raised in 2025, according to Home Office data cited in the joint investigation. The cohort where the algorithm performs worst is also the cohort most often processed.
Then came the procurement. In May 2026, the Home Office spent more than $400,000 on face-scanning technology from Cognitec, a German vendor, according to the UK's Contracts Finder procurement record. Cognitec was one of the seven algorithms tested in the 2025 trial. The leaked report does not name which of the seven was the best performer, so the public record cannot yet confirm that the Home Office bought the worst-tested vendor, the best, or some middle rung. What is on the public record is that the agency signed the contract without naming which algorithm its own test had singled out.
Lighthouse Reports then ran its own audit, using the US National Institute of Standards and Technology Face Recognition Vendor Test as a benchmark baseline. NIST's published results show that even the best-controlled algorithms carry roughly 2.5 years of mean absolute error in lab conditions. Lighthouse's audit, documented in its methodology page, found that Cognitec misclassified twice as many 16-year-olds as adults on lower-quality border-crossing photos as it did on higher-quality visa photos. Border photos are the photos the Home Office will actually be running.
The Home Office disbanded a scientific advisory committee on age estimation while exploring whether to deploy the AI. Tim Cole, an emeritus professor at UCL's Institute of Child Health and a named former member of that committee, called facial age estimation "hideously inaccurate" in comments to the joint investigation. He is on the record. The committee itself is not on the record, because it no longer meets. The agency's own guide for using FAE to support initial age decisions describes the system as "additional to human judgment." That phrasing is doing real work: it positions the algorithm as a screening aid rather than a decision-maker, which is the legal posture that lets a misclassification survive an appeal.
The "additional to human judgment" framing also defines the appeal window. Initial asylum age decisions are made inside a 72-hour border window, after which the consequences of being classed as an adult are already in motion. Since 2010, 40% of people facing age assessments have been classed as adults, per Home Office data cited in the investigation. Misclassification routes children into adult-only detention. The Home Office says that in cases of uncertainty, individuals will be treated as children pending further assessment. The structural problem is that the further assessment arrives after the initial decision has already shipped the consequences.
This is not a UK-only story. Facial age estimation is already mandatory or in trial for online age checks, for porn-age verification in several US states, for Australia's social media age ban, and for UK bar and shop regulatory sandboxes, as the broader age-checked internet rollout has documented. The deployment at the asylum border is the highest-stakes version, because there is no consumer who can close the tab and try a different service.
Sixty-two organizations, led by the UK civil-society group Foxglove, sent an open letter to the UK government asking the Home Office to scrap the FAE plan. Martha Dark, Foxglove's co-executive director, is on the record criticizing the deployment. The letter landed in the same week as the joint investigation. Policy is in flux: rollout is currently scheduled for 2027, and civil-society pressure combined with the published test results could delay or alter the program between now and then.
The thing worth watching next is not whether the algorithm improves. It is whether the agency that bought it rebuilds the kind of internal scientific review that could have flagged the failure mode before the contract was signed. The committee that could have done that work has already been disbanded. The contract has not.