The UK government signed a contract in June 2026 to deploy AI face-scanning technology at its border, using it to estimate whether asylum seekers who arrive without documents are adults or children. The technology scans a photograph and analyzes facial features to produce an estimated age range. The contract runs for three years and is worth roughly $430,000, with German firm Cognitec providing the technology through a British IT supplier called Akhter Computers.
This matters because the legal line is absolute. Under UK law, anyone under 18 arriving unaccompanied is classified as a child, placed in local authority care, given access to education, and afforded legal protections that adults simply do not receive. A face scan that adds a few years to someone's apparent age does not just produce an administrative inconvenience: it can strip a 14-year-old of child protections and put them in an adult detention facility.
The problem is that these systems do not work well at the boundary that matters. Independent US government benchmarks, run by a federal agency that tests this kind of technology, consistently show that accuracy is weakest in exactly the 16-to-18 age range. Errors are also larger for women and for people from regions that are underrepresented in the datasets used to train these systems.
The Home Office's own internal tests confirmed all of this before it signed the contract. The best system it tested showed errors averaging 4.6 years for young Sub-Saharan African women. That is not a marginal failure: a 13.5-year-old girl could be read as an 18-year-old adult. Sub-Saharan Africans are currently the largest group going through age assessments at the UK border.
There is a second layer to the problem. The internal tests were run mostly on good-quality photographs. Asylum seekers arriving after dangerous Channel crossings are photographed under difficult conditions, often exhausted and stressed from travel. The Home Office's own report noted that stress and physical hardship appeared to affect how old someone's face looks, and that photos taken at first encounter were consistently worse quality than follow-up photos. Worse photos mean larger errors.
Research on how people behave with AI decision-support tools adds another concern. Even when the system is officially just one input among several, officers working under time pressure tend to treat an algorithmic number as an anchor. A range becomes a number. A suggestion becomes a conclusion.
The Home Office disbanded an independent scientific committee that had been advising it on age assessment methods, reportedly while the AI program was still being developed. Former committee members say they were never given the chance to formally flag their concerns before the group was shut down.
To be fair, the current system is not working well either. Official data from July to December 2025 showed that 17 percent of people initially classified as adults at the border were later found to be children after a fuller assessment. The volume of age assessments has grown to over 6,000 a year, compared to between 300 and 1,200 a year in the previous decade. Human-only assessments have their own record of poor documentation and inconsistent practice.
But introducing a tool with known demographic errors into a process that already struggles, without clear training standards or published protocols for how officers should handle disagreements between their own judgment and the system's output, does not obviously make things better. The trial at Dover will produce data. What matters is whether that data is published, scrutinized, and acted on before the wider 2027 rollout rather than after.
For business operators, the immediate relevance is not the asylum question itself. It is the precedent. Face-scanning as a fast, cheap substitute for document-based verification is spreading: online age checks, bars, retail, borders. The same accuracy gaps that appear in government use will appear anywhere this technology is deployed, and the same tendency for humans to defer to an algorithmic number rather than question it will apply in every context.