Retrospective facial recognition runs at far greater scale than live, with less scrutiny

Live facial recognition gets the cameras, the protests and the court cases. Retrospective facial recognition, the version that runs after the fact against a database of custody photographs, gets almost none of that attention, despite running at far greater volume and, on the regulator's own findings, being governed less well.

The scale nobody sees

More than 25,000 retrospective facial recognition searches are run each month against the Police National Database, a gallery of more than 19 million custody images. Nothing about this happens in public. An officer takes an image from CCTV, a phone or a doorbell camera and searches it against the database from a desk. There is no van, no sign, no member of the public aware it is happening.

Live facial recognition, by contrast, requires a camera in a public place, a watchlist, and typically some form of public notice. It is visible by design. That visibility is precisely why it has been through the Court of Appeal and the High Court, while retrospective search, running at roughly the same order of magnitude in monthly volume as live deployment's annual arrest count, as recorded on the deployment tracker, has faced comparatively little scrutiny.

The regulator found the gap directly

The Information Commissioner's Office audited five forces between June 2025 and March 2026: South Wales and Gwent, Essex, Leicestershire, West Yorkshire and Greater Manchester. Across the programme, compliance was generally higher for live facial recognition than for retrospective searches. That finding runs against how the public debate is weighted.

Greater Manchester Police is the clearest illustration. The force received reasonable assurance of compliance for live facial recognition, but only limited assurance for retrospective use. The ICO attributed the lower rating to the absence of a policy governing the force's use of the Police National Database, and to a lack of defined roles for using the technology. West Yorkshire Police, audited in the same programme, received reasonable assurance for both. The difference between the two forces is not the technology. It is whether a policy existed.

A confirmed bias finding in the algorithm behind the database

In December 2025 the National Physical Laboratory published an accuracy and equitability evaluation of the Cognitec algorithm used for these searches. The ICO stated the result without qualification in August 2026: bias in the algorithm increases the likelihood of incorrect matches for people in some demographic groups.

The Home Office and the National Police Chiefs' Council have put mitigations in place, including staff training, oversight reporting and equality impact assessments, and plan to replace the algorithm. The ICO says it continues to engage with HMICFRS, the Home Office and the NPCC on that replacement, and has reserved the right to take further regulatory action if needed.

That sequence is worth recording accurately rather than only as criticism. An independent test found a demographic disparity, a regulator stated it in plain language rather than diplomatic hedging, and the operator committed to replacing the system. Not every finding on this site follows that pattern.

The gallery itself is the older, unresolved problem

Separately from the algorithm, there is the question of who is in the database at all. Custody images are taken on arrest, and arrest is not conviction. Images of people never charged, or charged and later acquitted, have historically remained in the collection, with the burden of requesting removal falling on the individual rather than happening automatically. Nineteen million images is a substantial fraction of the adult population of England and Wales, and every retrospective search runs against a gallery that includes a significant number of people who were never convicted of anything.

The legal position moved in 2026, in the direction of more searching, not less

Accounts describing UK facial recognition as operating with no statutory basis at all are now out of date, though not in a way that resolves the underlying concern. The Crime and Policing Act 2026 opened police access to DVLA driver records, including photographs. In March 2026 the House of Lords rejected an amendment that would have blocked facial recognition searching of the roughly 50 million image DVLA database, voting 123 to 40.

That is a categorical expansion rather than an incremental one. A custody image database contains people who have at least been arrested. A driving licence database contains people who have done nothing but apply to drive. If that gallery becomes searchable in practice, the great majority of adults become biometrically identifiable to police by default rather than by any individualised suspicion.

The broader legal framework has been deferred to a further Police Reform Bill. A Home Office consultation on facial recognition closed in February 2026 with a government response promised within twelve weeks. As of this writing it remains unpublished, more than six months later.

Why this matters more than the live facial recognition debate, not less

None of this is an argument that live facial recognition deserves less scrutiny. It is an observation that the version of this technology running at the greatest scale, against the largest and least consented-to gallery, invisible to the people it is used on, has received a smaller share of the legal challenges, the media coverage and, on the regulator's own account, the governance.

What is not established publicly is how many searches result in a correct identification, a wrongful one, or no match at all, and how many of the 19 million images belong to people never convicted. Until those figures exist, judging whether this system is proportionate to what it achieves is not possible from outside it.