Deciding whether an undocumented person who just crossed the Channel in a small boat is 17 or 19 is genuinely hard. It also matters enormously. A child gets safeguarding protections, age-appropriate accommodation, and a different legal track. An adult does not. Get it wrong in one direction and an adult slips into the system for minors. Get it wrong in the other and a frightened, traumatized child is treated as an adult, housed with strangers, and potentially detained alone.

So the question is not whether age assessment is difficult. It is. The question is whether you should hand that difficult, high-stakes call to an AI system that your own civil servants tested and found wanting. The UK Home Office looked at that question and decided the answer was yes.

What facial age estimation is meant to do

Facial Age Estimation, or FAE, is software that scans a photograph of someone's face and predicts their age within seconds. The Home Office wants to use it as an additional tool for immigration officers making initial age decisions about small-boat arrivals who turn up without identity documents. The stated goal, in the government's words, is to "crack down on fake claims by small boat arrivals posing as children." Trials are planned through 2026, with operational use at the border from 2027. In May 2026 the department awarded a contract to Cognitec Systems, a Dresden-based facial recognition company, to supply the algorithm.

The official framing is that FAE only assists; human officers still decide. That is the same reassurance attached to nearly every consequential AI deployment, and it deserves the same skepticism, especially when the assistance is wrong in a patterned, predictable way.

A leaked test the department fought to keep secret

Here is the part that moves this from "contested policy" to "documented hazard." The Home Office ran its own internal evaluation, the Facial Age Estimation Performance Test, produced in April 2025 by the civil servants running the biometrics programme. It tested seven commercially available algorithms, including ones ranked at the top by the US National Institute of Standards and Technology, against more than 2.5 million images.

When Lighthouse Reports requested that evaluation under freedom-of-information law, the Home Office refused to release it, arguing it needed a "safe space" to debate the issue "away from external interference and distraction." The report leaked anyway. In June 2026, The Independent, Lighthouse Reports, and WIRED published a joint investigation, "Asylum by Algorithm," built on that leaked document and on independent analysis of NIST benchmark data for the chosen vendor.

What the department's own test found is not a ringing endorsement. The "best performing" system tended to overestimate teenagers' ages: on average it predicted 17-year-olds to be 18 or older. It performed worse on female faces. And it showed "substantial deviations" for Sub-Saharan African faces, with an error rate roughly double that of other groups. For female Sub-Saharan Africans, the system's guesses were off by an average of 4.6 years, which means a girl of around 13 or 14 could be assessed as an adult.

A separate Lighthouse audit of public benchmark data for the government's chosen provider found the technology misclassified more than a third of 16-year-olds as adults, and in some tests gave the wrong assessment as often as 70% of the time. The report also noted that the few photos taken at first encounter, exactly the moment an officer would use FAE, were routinely worse quality than later ones, and that the stress and visible ageing of a long, traumatic journey can throw the systems off further.

Bias at exactly the boundary that matters

Layer these findings together and the failure is not random noise; it is concentrated precisely where it does the most damage.

The decision that matters in age assessment is almost always the 16-to-18 call. Nobody is using face-scanning AI to tell a toddler from a forty-year-old. The whole point is to adjudicate the borderline cases, and the borderline is where the Home Office admits the tool is weakest: even the "top systems" carry an error margin of around 2.5 years right at the 16-to-18 boundary, according to figures cited in the open letter from rights groups. A 2.5-year margin is catastrophic when the entire decision turns on whether someone is just under or just over 18.

Then add the demographic skew. The cohorts the tool is least accurate for, people from countries such as Eritrea and Sudan, are among the most common nationalities arriving by small boat. The population the Home Office most wants to run through FAE is substantially the population the technology is worst at reading. Critics have called this "baked-in racial bias," and on the department's own numbers that is not hyperbole; it is a description of the error distribution. Girls and people of color absorb the largest errors, and the consequence of an error is a child being treated as an adult.

Does a tool that isn't live yet belong here

Worth being precise about scope, because this site is careful about it. FAE has not yet been used to misclassify a real child in production; the rollout is slated for 2027, with trials during 2026. There is no confirmed individual victim to point to, and the broader argument about asylum policy is a political fight this catalog does not adjudicate.

What is documented, and what earns the headstone, is narrower and concrete: an AI system whose measured inaccuracy and demographic bias are recorded in the deploying agency's own leaked evaluation and in independent NIST-based analysis, being pushed toward deployment in a safeguarding context where the predictable failure mode is treating children as adults. That is a deployment hazard in the clearest sense. The harm is not speculative hand-wringing about what AI might someday do; it is the gap between what the testing showed and what the department decided to do anyway. Choosing to field a system you have already measured as unreliable, for the exact population it is least reliable on, is a decision that creates a foreseeable blast radius before a single border officer switches it on.

Disbanded advisers and missing assessments

The process around the decision does not inspire confidence either. Professor Tim Cole, an emeritus professor of medical statistics at University College London and a former member of the Home Office's own Age Estimation Science Advisory Committee, publicly questioned both the reliability of FAE and the government's decision to disband that advisory committee shortly before announcing the AI plan. Disbanding your scientific advisers and then proceeding is a particular sequence of events.

In June 2026, 62 organizations, coordinated by the non-profit Foxglove and including Amnesty International, Human Rights Watch, Liberty, the Electronic Frontier Foundation, and the Open Rights Group, sent an open letter to the border security and asylum minister urging the department to halt the rollout. They pointed out that the government had not published detailed test results, methodologies, an Equality Impact Assessment, or a Data Protection Impact Assessment that would let anyone independently scrutinize the claims of "promising" accuracy. As Foxglove's Martha Dark put it, children who have often suffered serious trauma "should not be the test subjects for experimental tech that has baked-in inaccuracy and racist bias," warning that errors could force vulnerable children, alone, into adult detention centres. The Home Office says it has asked the National Physical Laboratory to review testing as FAE is implemented, which is a reasonable thing to do before deployment, not during it.

What this comes down to

Strip away the politics and the engineering reality is simple. The Home Office commissioned a test, the test said the tool is least accurate at the only boundary that matters and most biased against the people it will most often be used on, the department tried to keep that test from the public, and it is proceeding to deployment regardless. Age assessment is hard enough that you might reasonably want a tool to help. But a tool that confidently rounds frightened teenagers up into adulthood, with the largest errors falling on girls and on children from a handful of African countries, is not help. It is a way to make a hard call faster and to be wrong about it in a manner that is now, thanks to the department's own files, entirely predictable.