The UK Home Office plans to use “facial age estimation” at the border in 2027: AI reads a photo of a face and estimates age to help officers make an initial judgment on whether an asylum seeker is a child or an adult. According to the guidance the Home Office published in May 2026, the technology is still being tested, won’t be relied on alone to make a definitive age decision, and “the decision on whether to treat an individual as a child or an adult will always be taken by the officer.”
In June, the investigative outlet Lighthouse Reports, with WIRED, published an analysis based on a leaked internal report and test data from the US National Institute of Standards and Technology (NIST): the best-performing vendor’s system tends to overestimate minors’ ages, on average placing 17-year-olds above 18; for female minors from Sub-Saharan Africa, the average error reaches 4.6 years. On NIST’s data, the system classified more than half of 16-year-old West Africans as adults, compared with under a quarter of 16-year-old Eastern Europeans.
“A person makes the final decision” is a common assurance. But it only holds if the person can actually overrule the AI: they need time, evidence, and authority, and the person being judged needs to know AI was involved and be able to appeal. When AI’s judgment changes how someone is treated, these must be designed before launch, not patched in later.
Overridable on paper, no time to override in practice
The investigation notes that age assessments usually happen at first encounter, when officers may be under heavy pressure, and the Home Office’s own report found that photos taken at first encounter are often of poorer quality. With many cases, little time, and a single number on screen, “assistance” easily becomes the actual decision. People can overrule it in name, but in practice lack the time, evidence, or authority to do so.
Errors aren’t shared evenly either. If some groups have much larger errors, “good overall accuracy” hides the risk those people bear. The investigation also describes a dilemma: raising the age threshold that triggers further assessment reduces minors treated as adults but increases adults treated as children, and widens the gaps between groups.
Answer four questions before launch
Purpose boundary: Does the AI result only flag “more information needed,” or does it directly change accommodation, protection, accounts, costs, or eligibility? If it changes them directly, people must lead the process.
Error distribution: Do the system’s errors differ across age, gender, ethnicity, and photo quality? Average accuracy alone isn’t enough.
Authority to overrule: Who can overrule the AI? What documents, interviews, or second opinions are needed, and how quickly is it handled? If frontline staff can only follow the number, it’s automated decision-making.
Records and appeals: Does the person being judged know AI was part of the process? Can they get the reasons and records and lodge an appeal? The Home Office guidance doesn’t describe a formal process for individuals to challenge an estimate.
The safer position is for AI to trigger human review rather than end it. The Home Office process already has a version of this: unless two officers independently judge that appearance “very strongly suggests” the person is significantly over 18, the person is treated as a child and referred to the local authority for a fuller assessment.
Small teams face the same problem
This isn’t only a government matter. A platform using AI to decide whether users are adults, an insurer or lender using AI to screen applications, a school using AI to detect cheating: the structure is the same. AI’s judgment changes how a person is treated. Whenever that’s true, answer the four questions above first, and don’t launch if you can’t.
For letting users return to the original process when AI gets it wrong, see How Fast to Roll Out AI Depends on Whether Users Can Back Out. For when AI should stop and hand off to a person, see Before Connecting AI to a Workflow, Define When It Should Stop and Ask a Person.
Four questions won’t make the model more accurate
These four questions give errors a way out, but they can’t make the model more accurate or judge whether a use should exist ethically or legally. Decisions involving immigration, child protection, healthcare, or financial eligibility need legal and domain experts, not just the team adopting the AI. The Home Office says the UK’s National Physical Laboratory is conducting an independent review; until it’s published, the technology’s real-world performance remains to be confirmed.
AI handoff card
We plan to use AI to help judge something that affects people (such as age, identity, eligibility, or rule violations), and want to make sure errors have a way out first.
The method comes from this article: https://boosterminiclass.com/en/posts/facial-age-estimation-human-review-checklist/
Its view: "a person makes the final decision" only counts if the person can actually overrule it. When AI's judgment changes how someone is treated, design these before launch: the purpose boundary, how errors are distributed across groups, who can overrule and with what evidence, and whether the affected person knows AI was involved and can appeal. Let AI trigger human review rather than end it.
First ask me: what the AI will judge, what the result changes for the affected person, and who makes the final decision in the current process.
Then:
1. List who loses protection, services, accounts, or opportunities when the AI is wrong, and the consequences of errors in each direction;
2. List the error data to request from the vendor: error rates broken down by age, gender, ethnicity, and data quality;
3. Design the override process: who can overrule, what evidence is needed, how quickly it's handled, and whether frontline staff have enough time and authority;
4. Design notification and appeals: how affected people learn AI was involved, which records they can get, and whom they can appeal to.
Mark anything that can't be answered as "must be completed before launch," and don't judge for me whether this use is legal.
References
- UK Home Office official guidance: Facial Age Estimation using AI to support initial age decisions: a guide — https://www.gov.uk/government/publications/facial-age-estimation/facial-age-estimation-using-ai-to-support-initial-age-decisions-a-guide-accessible (2026-05-29)
- Lighthouse Reports: How we analysed AI used to guess asylum seekers’ ages — https://www.lighthousereports.com/methodology/how-we-analysed-ai-used-to-guess-asylum-seekers-ages/ (2026-06-18)
- NIST official testing: Face Recognition Vendor Test — Age Estimation — https://pages.nist.gov/frvt/html/frvt_age_estimation.html [accessed: 2026-10-11]
