A Tool That Decides Who Gets Child Protection
On June 1, 2026, the UK Home Office activated a three-year, £322,000 contract with Akhter Computers and German facial recognition firm Cognitec to pilot AI facial age estimation (FAE) at Dover. The system photographs an asylum seeker who has arrived without identity documents and produces an estimated age. That number feeds directly into a decision that is anything but administrative: under UK law, anyone deemed to be under 18 receives placement in local authority care, access to education, and a suite of legal protections that adults at the border do not. The technology's own guidance, published by the Home Office, concedes it has an error margin of around 2.5 years at the 16-to-18 boundary — the only threshold being assessed.
The Government's Case Deserves a Hearing
Before examining the system's failures, the strongest argument for deploying it deserves honest engagement. Undocumented age claims are genuinely hard to adjudicate. Human caseworkers assessing unfamiliar faces — without documentation, without bone-density scans, in high-pressure processing environments — typically err by around eight years, considerably worse than leading algorithms under controlled conditions. The Home Office cites evidence that some adults misrepresent their age to access child-specific protections. Minister Alex Norris stated plainly: "For too long, adult migrants making false age claims have exploited the system." An additional, bounded data point that can be logged, audited, and compared against officer judgment is, in principle, a legitimate supplementary tool.
The Home Office also included formal safeguards in its guidance: immigration officers retain final decision-making authority, two independent assessors must agree before treating someone as "significantly over 18," and social worker assessments must be accounted for. These commitments are not trivial. And Home Office data collected before the AI rollout already showed 326 people initially assessed as adults were subsequently reclassified as children — confirming that human-only assessment is also error-prone.
Error Rates Are Highest for the Populations Being Assessed
The problem is empirical, not ideological. The Home Office's own published guidance acknowledges that "error rates were almost always higher for female faces" and that "FAE performance can vary depending on ethnicity, skin tone, gender, place of birth and quality of input image." These are not edge cases — they describe the majority of the population arriving at Dover via small boats and lorry crossings, which skew heavily toward nationals of West and Sub-Saharan African countries.
Investigative outlet Lighthouse Reports, drawing on NIST benchmark data, found that more than half of 16-year-old West Africans were misclassified as over 18, compared to fewer than a quarter of Eastern Europeans. A leaked Home Office evaluation of seven algorithms — tested against more than 2.5 million images — found Cognitec's best performer worked worst on Sub-Saharan Africans: among girls from that region, the average age estimation error reached 4.6 years. A 14-year-old girl from Sub-Saharan Africa could algorithmically register as an adult, lose her child protections, and be channelled toward removal — with an officer's "final say" shaped by an estimate the officer has no independent means to interrogate.
There is a further compounding factor the guidance does not address: the physiological effects of the asylum journey itself. Severe trauma, malnutrition, chronic dehydration, and sleep deprivation can visibly age a child's face. Systems trained on controlled photo datasets have no mechanism to discount those effects, and the Home Office's guidance offers no protocol for flagging cases where such factors may be present.
The Accountability Gap Is the Actual Failure
On June 18, 2026, 62 organisations — including Amnesty International, Human Rights Watch, Liberty, the Open Rights Group, Foxglove, and the Electronic Frontier Foundation — wrote to Minister Norris demanding an immediate halt. Their four concerns form a coherent indictment: discriminatory performance that maps precisely onto the demographic profile of those being screened; an error margin the Home Office has itself documented at the one age boundary in question; no published legal basis for collecting and processing photographs of children to train the system; and an almost total absence of public accountability documentation.
The Home Office claims "extensive testing" across diverse demographic groups has been completed and that results are "promising." But neither the results nor the methodology have been published. No Equality Impact Assessment has been released. No Data Protection Impact Assessment has been made available.
"The Home Office has not published the testing results, methodologies, or impact assessments that would allow independent review of its claims about accuracy."
Under UK data protection law — the UK GDPR and the Data Protection Act 2018 — a high-risk processing activity involving biometric data from a vulnerable population requires a DPIA before deployment. The Equality Act 2010 similarly requires documented impact assessment for any policy likely to produce differential outcomes by race or sex. Both legal instruments apply here. Both appear to have been bypassed.
What Proportionate Deployment Actually Requires
The objection here is not that AI has no role in age assessment. A well-validated, demographically representative system with explicit confidence intervals, clear human-override protocols, and documented training-data provenance could be a proportionate supplementary tool. The Home Office's safeguards — human final decision, dual-assessor rule, benefit of the doubt — describe the right architecture.
But the architecture does not substitute for the evidence. Proportionate deployment requires: published demographic accuracy data broken down by gender, ethnicity, and region of origin; completed and public DPIA and Equality Impact Assessments; documented anti-automation-bias procedures; and an accessible avenue for individuals to contest an estimate before it determines how they are processed.
The government has the legitimate aim. It has yet to demonstrate the legal and evidential foundations required to pursue it with this system, on this population, at this scale.