When the UK Home Office published its guidance on Facial Age Estimation in May 2026, it included an admission that should have triggered an immediate halt: the AI system carries an error margin of approximately 2.5 years at the 16-to-18 age threshold — the precise range where a mistake determines whether a vulnerable young person receives child protection or is placed in adult detention.
On June 19, 2026, that admission prompted a coalition of 62 organisations — including EFF, Foxglove, and Human Rights Watch — to write to Alex Norris, Minister of State for Border Security and Asylum, demanding a pause on the planned 2027 deployment. The letter raised three core concerns: unpublished test results, the absence of Equality Impact and Data Protection Impact Assessments, and well-documented bias against women and people of colour.
The Case for the Technology
The government's rationale deserves fair hearing. Human age estimators are substantially worse: an immigration officer relying on a passport photo is typically off by around eight years, according to academic analysis by Loughborough University researchers. A 2.5-year margin for an AI system represents, in the aggregate, a real improvement. Fraudulent age claims are a documented problem in the UK asylum system, and the Independent Chief Inspector of Borders and Immigration launched a formal inspection of existing age assessment processes in September 2024, finding inconsistent training and patchy data quality across intake units.
The cost argument is also real. The contract awarded to Akhter Computers — working with German facial recognition specialist Cognitec — is worth £322,000 across three years, a fraction of the cost of alternative assessment methods. The Home Office frames FAE as a "supplementary tool" to inform, not replace, human officers; its May 2026 guidance is explicit that the technology "is not currently in operational use" and will serve only to provide additional information.
Minister Norris stated the ambition plainly: "We are rolling out AI technology to put a stop to this, ensuring those who game the system are identified, detained and removed without delay."
Where the Maths Breaks Down
The problem is that a 2.5-year margin is not evenly distributed. It concentrates at exactly the decision point that determines legal status. At the 16-to-18 boundary, NIST benchmarks show error rates for leading systems are materially higher than the overall average. The Home Office's own guidance acknowledges this without explaining how officers should account for it in practice.
Data published by Lighthouse Reports, drawing on NIST benchmarks and leaked Home Office internal test results covering seven competing vendors, is considerably more specific — and more alarming. Cognitec's system, the contracted vendor, misclassified over two-thirds of 16-year-olds as adults when applied to US border photos. Among 16-year-old West Africans specifically, the system predicted more than half as being over 18; among Eastern Europeans the same age, the figure was under a quarter.
The Home Office's own internal evaluation found that the best-performing vendor still systematically overestimated teenagers' ages. For sub-Saharan African girls, the average error reached 4.6 years — meaning a 14-year-old from that demographic could routinely be assessed as an adult. The government guidance itself concedes that performance "varies depending on ethnicity, skin tone, gender, place of birth and quality of input image," and that error rates were "almost always higher for female faces."
That the system will primarily be applied against populations arriving from sub-Saharan Africa, the Middle East, and Central Asia — exactly the groups where bias is worst — is not incidental. It is the use case.
The Transparency Deficit
Beyond accuracy, the coalition letter identified a governance failure: the Home Office has not published its test results, Equality Impact Assessment, or Data Protection Impact Assessment. For a technology processing biometric data — classified as special category data under UK GDPR, subject to stricter necessity and proportionality tests — this is not procedural paperwork. It is a legal precondition.
Without a published DPIA, there is no public record that the processing meets UK GDPR requirements. Without an Equality Impact Assessment, there is no demonstration that the Equality Act 2010 has been applied to a system the Home Office itself acknowledges produces differential error rates by race and gender. The coalition was given 21 days to receive a response.
As Martha Dark, Foxglove's co-executive director, put it: "Errors by these tools could have serious consequences: vulnerable children being forced, alone, into adult detention centres."
The Asymmetry of Stakes
The legal consequences of misclassification run in one direction. Under UK law, unaccompanied asylum seekers under 18 receive local authority care, access to education, and specific safeguarding protections that adults do not. A child wrongly assessed as an adult at the border does not receive a corrective review in the same pipeline — they enter a different system entirely.
The ICIBI's 2024 inspection found that existing age assessment already suffers from "a culture of disbelief, racial bias, poor data systems, and lack of coordination with safeguarding professionals." Introducing a biased algorithmic signal into that process is unlikely to correct those pathologies. It is more likely to launder existing human bias behind a veneer of technical objectivity.
Anna Bacciarelli, Human Rights Watch's senior AI researcher, summarised the objection precisely: "Rolling out this experimental technology in such a critical and high-stakes use case is simply too risky."
What Responsible Deployment Would Require
AI in border processing is not inherently wrong. The technology will mature, bias correction methods will improve, and the aggregate accuracy advantage over human estimation is real at most ages. The question is whether deployment should precede validation, or follow it.
Three conditions would need to be met before this system could be considered fit for purpose: validated performance specifically on populations arriving at UK borders rather than US visa or NIST benchmark datasets; full public disclosure of Equality Impact and Data Protection Impact Assessments; and genuine human oversight with officers trained to understand when confidence intervals make the system's output uninformative.
None of those conditions have been met. A £322,000 contract has been signed and a 2027 deployment date has been set. The Home Office's own guidance concedes the technology "has no way to precisely determine someone's age." Deploying it anyway — with unpublished results and unresolved bias — is not a border security policy. It is an experiment conducted on the people least positioned to challenge the results.