Estimating the age of a child

On 29 May 2026, the Home Office published its accessible guide to Facial Age Estimation; the AI tool it plans to test through 2026 and deploy at the UK border in 2027 to support initial age decisions on age-disputed asylum claimants.

Most of the commentary so far has focused on whether AI can estimate a child's age accurately. That is the wrong starting question. The right starting question for any public lawyer reading the guidance is whether the proposed deployment is consistent with the legal architecture that already governs decisions of this kind. That architecture is more substantial than the published guidance acknowledges, and it has tightened materially over the last twelve months.

This piece sets out the five legal anchors that bear directly on FAE deployment. Each is supported by current authority. Taken together, they suggest that the technology as presented in the 29 May guidance sits closer to the edge of legal defensibility than the Home Office's confident framing implies.

The headline points from the 29 May guidance are these. FAE will be used to support the immigration officer's initial age decision at the point of first encounter; typically a small-boat arrival. The technology will not be used to automate the decision. The officer remains the decision-maker. A more comprehensive Merton-compliant assessment by a local authority remains available downstream, where there is residual doubt. The guidance is, in tone, reassuring.

Among the technical detail, on the Home Office's own data, is this:

"Algorithms also tend to perform better for age threshold determinations than for estimating specific age. Such precise estimations are less accurate at the important 16 to 18-year-old boundary. Even the top systems have an error margin of around 2.5 years here."

And:

"There is evidence in testing data that FAE performance can vary depending on ethnicity, skin tone, gender, place of birth and quality of input image. NIST found that error rates were almost always higher for female faces, although it didn't find out why as testing was purely on performance rather than how algorithms work."

These are not anti-AI campaigners' claims. These are the Home Office's own concessions, on the record, in published guidance. They establish two things as a matter of public knowledge from 29 May 2026 onwards: first, that the technology is least accurate at the precise age band where the legal consequences are most acute; and second, that performance varies by protected characteristic. After this date, no public authority deploying FAE in this domain can plead that those risks were unforeseeable.

That is where the public law architecture begins to engage.

The 29 May guidance opens by reminding readers that the Home Office has "a legal duty to safeguard and promote the welfare of children when carrying out immigration functions." That is the duty under section 55 of the Borders, Citizenship and Immigration Act 2009, a duty that the courts have treated, in a long line of authority going back to ZH (Tanzania), as substantive rather than aspirational.

The framing the Home Office has chosen that FAE will "strengthen the age assessment process" and "improve the safeguarding of children" is the section 55 framing. It puts the whole deployment inside the welfare duty. That is a strong rhetorical move; it is also a legal commitment. If FAE in practice produces a higher rate of children being initially treated as adults; children diverted, by definition, into adult asylum accommodation, away from local authority safeguarding, away from the protections that flow from the Children Act 1989, the deployment will not have strengthened the age assessment process. It will have weakened it.

The published 2.5-year margin of error at the 16-to-18 boundary is, on the Home Office's own data, an error rate that materially increases the risk of that outcome. A child of 17 estimated by FAE at 19.5 is, on the face of the system's own performance, the predictable case. Section 55 is engaged the moment the deployment is designed.

The duty articulated in Secretary of State for Education and Science v Tameside MBC [1977] AC 1014 requires a public authority to take reasonable steps to acquaint itself with the relevant information to enable it to discharge its function. That duty does not relax when an AI tool is in the room. It tightens.

Lord Sales, in his keynote lecture to the Government Legal Department's Annual Conference on 5 November 2025, treated the Tameside duty as capable of generating a requirement that the public authority acquaint itself with the performance, accuracy and known failure modes of the ADM system it deploys. After 29 May 2026, the Home Office cannot rely on opacity. The performance data, including the differential accuracy at the 16-18 boundary and the variation by ethnicity, is published. The Tameside duty therefore requires the individual officer to understand what the tool's estimate can and cannot tell them in the case in front of them.

The officer who says "the system says 19" without being able to articulate the system's known error margin at that age band, the differential performance across protected characteristics, or the way the estimate should be weighed against the officer's own observations, has not discharged the duty. The published guidance by putting those limitations on the record has made the Tameside test easier to fail, not harder.

The 29 May guidance repeats, in several places, that FAE will not replace the officer's judgement: "It allows officers to test their judgement against an estimate with known limits to its accuracy."

That framing is reassuring on paper. In operational reality, it raises a public law question that Lord Sales identified directly in the November 2025 lecture. The risk Lord Sales described adopting Mireille Hildebrandt's concept of "judgemental atrophy" is precisely the risk that human decision-makers, presented with ADM outputs under workload pressure, progressively cede their judgement to the system. The risk is not the rare officer who consciously defers; the risk is the rhythm of high-volume practice under pressure, in which the system's output becomes the path of least resistance.

The 29 May guidance describes the operational context in language that maps directly onto Lord Sales's concern: "immigration officers working in challenging circumstances, often under pressure to quickly process lots of new arrivals." That is the environment in which judgemental atrophy is most likely to occur. The common law principle against fettering of discretion articulated in British Oxygen Co Ltd v Minister of Technology [1971] AC 610 is engaged whenever, in practice, an algorithmic system is treated as determinative, even where the formal governance documents say it is only supportive.

The defensive question for the Home Office is whether the operational record will show, for the case in front of the court, that the officer actually exercised independent judgement; what the officer changed, what was overridden, why. If the record shows ratification rather than decision, the fettering argument is live.

The Public Sector Equality Duty under section 149 of the Equality Act 2010 requires the decision-maker to have due regard, in substance and contemporaneously, to the equality impact of policy and operational decisions. The leading authority on the PSED in the AI context is R (Bridges) v South Wales Police [2020] EWCA Civ 1058.

Bridges is the single most important precedent for any UK public authority procuring AI. The Court of Appeal held that South Wales Police had breached the PSED by failing to take reasonable steps to enquire whether its automated facial recognition technology was biased on grounds of race or sex even though there was no evidence that the technology was, in fact, biased. The duty of enquiry was triggered by the foreseeable risk of disparate impact. The absence of proof of actual bias was not a defence.

That standard enquiry, not proof applies with force to FAE. The Home Office's own 29 May guidance acknowledges that "FAE performance can vary depending on ethnicity, skin tone, gender, place of birth and quality of input image." That is the foreseeable risk. The PSED duty therefore requires the Home Office to demonstrate, contemporaneously and substantively, that it has interrogated the disparate-impact evidence, considered the demographic composition of the population the tool will be applied to, and put in place mitigations addressing the differential performance the technology vendors have themselves disclosed.

A generic statement that the Home Office is committed to equality, or that vendors take bias seriously, will not discharge the duty. Bridges puts that question beyond doubt.

The Working Together to Safeguard Children 2026 statutory guidance, published in March 2026, strengthens this analytical position further. The 2026 guidance imposes a positive duty on public authorities engaged in safeguarding work to identify and analyse patterns of racism and disproportionality. The Home Office is not within the strict scope of Working Together 2026; that guidance applies to children's services. But the cross-government policy expectation that disproportionality will be analysed and addressed, particularly in safeguarding contexts, is now a published feature of the regulatory landscape. A border-stage tool whose published performance data shows differential accuracy by ethnicity and gender sits awkwardly against that expectation.

The Home Office's strongest defensive argument, on the face of the 29 May guidance, is that the initial age decision is not determinative; that a Merton-compliant assessment by a local authority remains available downstream for any case where doubt persists. That argument has some weight. It is not, however, a complete answer.

The Merton case-law, beginning with R (B) v London Borough of Merton [2003] EWHC 1689 (Admin), establishes a standard for the holistic, welfare-led age assessment that local authorities must conduct when their statutory duties under sections 17 and 20 of the Children Act 1989 are engaged. Merton sits downstream. The 29 May guidance proposes to deploy FAE upstream, at the point of first encounter, to inform the "significantly over 18" two-officer threshold.

That upstream position matters. A young person who is treated, at the border, as significantly over 18 is by definition diverted into adult asylum accommodation and away from local authority children's services. They will not, in many cases, be referred for a Merton-compliant assessment at all, because the operational route to a local authority assessment depends on the initial age decision having gone the other way. The downstream Merton safeguard exists, on paper. In operational reality, whether a young person reaches it depends on the upstream decision the FAE tool is being designed to inform.

That brings the Article 8 ECHR analysis into focus. Article 8 protects respect for private and family life. Bridges held that South Wales Police's deployment of AFR was not "in accordance with the law" for Article 8(2) purposes because the legal framework left too much discretion and provided insufficient constraint. A border-stage AI tool that informs a decision channeling a young person into adult accommodation with potentially years of consequences for their access to safeguarding, education and mental health support is plainly an interference with Article 8 rights. The question is whether the published framework governing its use, the guidance, the operational rules, and the published constraints, is sufficient to satisfy the "in accordance with the law" requirement. An accessible explanatory FAQ on GOV.UK is not, on the Bridges standard, a published legal framework. Bridges set the bar higher than that.

The Data (Use and Access) Act 2025 came into force in its relevant provisions on 5 February 2026. Section 80 of the DUAA replaced Article 22 UK GDPR with Articles 22A to 22D. Article 22A defines a decision as "based solely on automated processing" where there is no meaningful human involvement. Article 22B prohibits solely automated significant decisions based on special category data unless explicit consent or specific legal authorisation applies.

Initial age decisions are significant decisions in the Article 22A sense; they produce a legal effect for the individual. The data involved is paradigmatically special category: it includes ethnicity data (the FAE algorithm's input is the face), and the decision is taken in the asylum context where data about religion, country of origin and family circumstances is engaged. Article 22B is therefore on the table.

The Home Office's defensive position will be that the immigration officer remains the decision-maker, so Article 22A is not crossed. That is the same defensive position vendors of AI transcription tools in social care take, and Lord Sales addressed it directly in November 2025: the question is whether the human involvement is meaningful in substance, not whether it is nominal in form. If the operational reality is an officer under workload pressure approving the FAE estimate after a brief review, the Article 22A meaningful human involvement question is engaged. If the threshold is crossed, Article 22B's near-prohibition follows.

None of this is an argument that FAE cannot lawfully be deployed. The argument is narrower and more useful: the deployment that the 29 May guidance describes has not yet been shown to meet the legal architecture that applies to it. Six anchors; section 55, Tameside, fettering, the PSED following Bridges, Merton and Article 8, and the post-DUAA Article 22 framework are all engaged, on the face of the published guidance.

For the deployment to withstand legal scrutiny, the Home Office will need to show, at minimum:

• A published legal framework, not an accessible explainer, specifying the permitted use of FAE, the operational constraints, the supervision arrangements, and the review processes, sufficient to meet the Bridges "in accordance with the law" standard.
• A substantive PSED analysis, completed before deployment, engaging in detail with the differential performance data the Home Office has itself disclosed.
• An evidential audit trail, by design, capable of showing in each case what the officer changed, queried or overrode in light of the FAE estimate — the Article 22A and Lord Sales answer to fettering.
• A meaningful operational route to a Merton-compliant assessment for any young person whose claimed age sits within the published margin of error of the FAE estimate, particularly where the estimate sits at or just over the 18 boundary.
• An equality impact assessment that engages with the published performance variation by ethnicity, skin tone, gender and place of birth, and articulates the mitigations the Home Office has put in place.

The 29 May guidance is not, on its face, that document. It is a public-facing explainer with a confident tone. The operational governance framework that meets the legal architecture is yet to be published. Until it is, the Home Office's confidence is doing a great deal of work.

If FAE deployment is to be defensible, the architecture described above is the architecture against which it must be built. The 2026 trial year is the moment when that architecture should be developed, published and tested. The risk, otherwise, is that 2027 begins with a tool in operational use and a legal framework that the courts will be asked to assess on the back of the first contested case. That is precisely the position South Wales Police found itself in. The outcome there is on the record.