A Tool, Not a Verdict — but the Stakes Are Binary
On 1 June 2026, the Home Office activated a £322,000 contract with Akhter Computers Limited, subcontracting the core algorithm to Cognitec Systems, a Dresden-based biometrics company. The system uses machine learning trained on millions of age-labelled photographs to generate a numerical age estimate from a face image taken at the border. The department has been explicit about its framing: immigration officers retain final authority, and cases where doubt persists are referred to local authorities for a full Merton-compliant assessment by qualified social workers. Testing will continue throughout 2026, with operational deployment planned for 2027.
That framing is the strongest version of the government's case, and it deserves to be heard. The tool is not designed to be a judge. It provides an additional data point in a decision that would otherwise rest entirely on an officer's visual impression of a person who has arrived exhausted, possibly malnourished, and potentially traumatised. The Home Office's published guidance acknowledges that under the existing human-only process, 326 individuals initially classified as adults were later determined to be children — real harm with real consequences. The problem the technology is trying to solve is genuine.
Where the Numbers Break Down
The difficulty lies not in the intent but in the performance gaps the government's own published guidance acknowledges — and the demographic pattern those gaps follow.
The Home Office guide, published 29 May 2026 on GOV.UK, cites National Institute of Standards and Technology (NIST) testing showing the technology's mean absolute error has improved from 4.3 years to 3.1 years on visa-quality photographs. That is measurable progress. But the same guidance acknowledges that "even the top systems have an error margin of around 2.5 years" at the critical 16–18 age boundary — precisely where the system will be applied.
A ±2.5-year error centred on a 17-year-old produces estimates ranging from 14.5 to 19.5. That span encompasses both "child requiring full statutory protection" and "adult processed through the standard asylum route." The mean-average figure, in other words, obscures the operational risk.
The aggregate error rate also conceals a structural disparity. According to an investigation by Lighthouse Reports, drawing on internal Home Office evaluation documents, more than half of 16-year-old West Africans tested were predicted to be over 18, compared with fewer than a quarter of Eastern Europeans. For Sub-Saharan girls specifically, the average error reached 4.6 years — meaning a 14-year-old could routinely receive an estimate placing her well into adulthood. The Home Office's own guidance concedes that "FAE performance can vary 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." These are not allegations levelled by campaigners — they are concessions published by the procuring department in a document intended to reassure the public about the technology's safeguards.
The Deployment Mismatch
The asylum-seeking population at the UK border is not demographically representative of the NIST test datasets on which headline accuracy figures are computed. Arrivals skew heavily toward exactly the cohorts for which the tested systems performed worst. Additional factors compound the problem: the Home Office's internal evaluation documents, referenced in Lighthouse Reports' investigation, note that trauma, malnutrition, and the physical stress of dangerous journeys may cause asylum-seeking children to appear older than their chronological age — noise the algorithm cannot distinguish from genuine ageing.
On 18 June 2026, Human Rights Watch and Foxglove published an open letter co-signed by 61 additional civil society organisations — 63 in total, including Amnesty International, Liberty, the Electronic Frontier Foundation, and the Open Rights Group — calling for an immediate halt. The letter cited what HRW termed "baked-in racial bias" in the evaluated systems and demanded that the Home Office publish details on "accuracy, efficacy, and discriminatory risks" before any operational rollout. As of publication, no Data Protection Impact Assessment or Equality Impact Assessment for the programme has been made public.
The absence of a DPIA is not merely an administrative gap. Under the UK GDPR and the Data Protection Act 2018, processing biometric data to derive actionable conclusions about individuals triggers mandatory DPIA requirements when the processing is "likely to result in high risk." Applying automated biometric processing to determine whether a person qualifies as a child under immigration law meets that threshold on any reasonable reading of ICO guidance on biometric recognition. That the department has not published one before committing to a three-year commercial contract is a sequencing problem that makes independent scrutiny impossible.
The Proportionate Path
The goal of a more consistent age-assessment process is legitimate, and the technology may yet improve enough to be responsibly deployed. But two conditions should be non-negotiable prerequisites for the 2027 rollout.
First, published disaggregated testing data. The Home Office should release its full evaluation results broken down by sex, age band, and region of origin, so independent researchers can verify the error rates used to justify the procurement. Without public data, no external assessment of the safeguards is possible.
Second, binding accuracy thresholds by demographic subgroup. An overall MAE of 3.1 years is meaningless if the worst-performing sub-group shows 4.6-year errors. The procurement contract should specify minimum accuracy standards that must be met across all demographic groups tested — not just in aggregate — as a binary go/no-go condition for deployment.
A three-year contract creates institutional momentum toward use. The asymmetry matters: if a 14-year-old is misclassified as an adult, the cost falls entirely on her. The cost of delaying deployment until the testing framework is robust falls on a budget line. The Home Office should resist the pressure to justify its contract spend before the accountability infrastructure — disaggregated accuracy data, independent equality assessment, published DPIA, demographic-specific accuracy floors — gives that deployment a legitimate basis.