UK starts face-age tests for asylum seekers next year, despite flawed AI error rates

Starting next year, the British government will use facial age estimation at the UK border, scanning asylum seekers' faces to predict how old they are. The point is simple: many arrivals will not have documents proving their age. The consequence is not. If children are incorrectly classed as adults, they can lose some legal protections and be placed in adult-only detention centers.

The catch, and the reason this is turning into a political and legal flashpoint, is that the underlying technology appears unreliable in exactly the way you do not want it to be. An investigation by WIRED and Lighthouse Reports, in collaboration with The Independent, obtained an internal UK government report detailing tests of facial age estimation technologies. Those tests show the systems regularly mistake children for adults and appear to contain serious bias problems, directly impacting the largest group of migrants subject to age assessments in 2025, according to data from the Home Office.

To understand why this matters beyond border policy, zoom out to how age verification has become a recurring requirement online. Age checks are already a daily reality for many people, from social media bans in Australia to porn restrictions in half of US states. That is the market context: when regulators, platforms, and content providers can point to “age verification” as a compliance tool, the ecosystem builds around it. AI-based approaches then become tempting because they look scalable. Move fast, reduce manual review, and centralize the decision in an automated pipeline.

Now the UK is trying to carry that playbook from the internet to the offline world. Facial age estimation works by taking a face image and predicting a likely age. In low-stakes settings, that kind of guess might be tolerable as a rough filter. At a border, it is a gatekeeper signal that can determine legal categories and detention placement. The internal report obtained by WIRED and Lighthouse Reports is therefore not just a technical critique. It is a warning about error costs, where the “wrong” classification can mean the difference between being treated as a child or being treated as an adult.

This is also why the bias angle is especially damaging. The investigation reports that the systems appear to contain serious bias problems. Bias is not an abstract problem when the output is tied to legal protections. If the model systematically overestimates age for certain groups, it will not merely degrade accuracy metrics. It will shift outcomes, and that is the part regulators and boards should worry about, because it can turn a compliance process into a distributional injustice machine. When the stakes are detention conditions and legal status, you cannot treat bias as a statistical nuisance.

There is an additional policy tension: governments tend to face pressure to solve the “lack of documentation” problem. The source is clear that many asylum seekers arriving in the UK will not have documents proving their age. But this creates a decision trap. When you have a class of people who cannot produce paperwork, you are incentivized to look for alternative evidence. If an AI system is offered as “evidence,” then every misclassification becomes an operational risk and a public trust risk at the same time.

The investigation explicitly raises questions about both the effectiveness of the technology and whether it should be deployed in such high-stakes scenarios. That is the real strategic stake for decision-makers in similar roles, including lawmakers, platform policy teams, and any organization building age or identity enforcement systems. Even if an AI tool performs “well enough” in controlled tests, what matters is the error pattern at the boundary between child and adult. In this case, the internal tests described in the report show misclassifications that are consequential, and bias concerns that could compound those failures. When the largest group of migrants subject to age assessments in 2025 is at risk, “good intentions” will not fix the math.