A 30-month margin of error: the AI age-check the UK is rolling out on child refugees

The UK Home Office plans to scan the faces of refugee children with an AI whose own admitted margin of error is 30 months. Home Office figures obtained by the Helen Bamber Foundation show 755 children were misclassified as adults at the border in 2025, before the new tool was switched on.

In May 2026 the Home Office awarded the contract for that tool to Cognitec, a German vendor that bills itself as "the face-recognition company". The system reads a face from a camera at a border post and returns a numerical age estimate, which an immigration officer is supposed to weigh alongside the existing process for initial age decisions. The Home Office's public framing is that the AI assists, rather than replaces, human judgement.

A June 2026 investigation by Lighthouse Reports, a Dutch investigative newsroom, found that Cognitec's system systematically over-predicts the ages of sub-Saharan African minors, meaning the tool leans toward calling children older than they are. The underlying data and methodology are public on GitHub. The direction of the error matters: a child the system reads as an adult can be detained in adult accommodation, removed under adult removal procedures, or sent to housing that has previously been the site of protest-related violence. An adult the system reads as a child is, at worst, given extra protection.

The Home Office has acknowledged that even the best age-estimation systems carry a 30-month margin of error. It has not published the confidence interval, age band, or demographic slice the 30-month figure refers to, or whether the same margin holds for darker-skinned faces. On a screening whose entire purpose is to separate children from adults, a 30-month error band is large enough to make the call depend on the system's bias rather than the child's actual age.

Refugee-sector groups say the 30-month figure is not a uniform technical fact, and that it is shaped by the population it is run on. Maddie Harris of the Humans for Rights Network told The Guardian that the AI rollout amounts to a "cheap process" that will not improve existing misidentification, and warned of an "adultification" effect for children from countries such as Sudan and Somalia. The Refugee and Migrant Children's Consortium has flagged that trauma can physically age a child's appearance, compounding the bias that the Lighthouse audit identifies. Sixty civil-society groups, coordinated by the UK tech monitor Foxglove, have asked the Home Office to drop the technology, characterising it as "racist bias" rather than a calibration problem.

The Home Office's defence is that the AI is layered onto officer judgement, not substituted for it. That structure is the point of the critique: if the underlying process already misclassified 755 children in a single year, an automated read that is, on the government's own admission, off by up to 30 months is being added to a process that was not accurate to begin with. The Lighthouse finding makes the margin a directional error: against sub-Saharan African children, the tool is more likely to call them older, not younger.

The Home Office has not published a deployment date for the new screening layer.