Live facial recognition tends to attract most of the public attention, with its vans, signage and court challenges, yet it is the retrospective kind that does most of the day-to-day work. Forces in England and Wales ran 291,341 retrospective searches in the year to 31 March 2026, each one comparing a still image from CCTV, a doorbell camera or a mobile phone with the custody photographs held on the Police National Database.
On 12 November the policing inspectorate, HMICFRS, published the first national inspection of how that work is carried out. The Home Secretary commissioned it a year ago, and inspectors visited seven forces and the National Crime Agency during March and April: the Metropolitan Police, British Transport Police, Cheshire, Dorset, Gwent, Northumbria and Surrey.
Most of the coverage concentrated on the risk of wrongful arrest and on custody photographs that should have been deleted. Those findings are in the report and deserve the attention they received, but the report as a whole reads less like a criticism of the technology than a description of a tool that has grown faster than the record-keeping around it. That is a more ordinary problem than the headlines suggested, and in many ways a more fixable one.
A tool the inspectorate wants forces to keep using
It is worth being clear at the outset that HMICFRS supports the use of retrospective facial recognition. The report describes it as a valuable and proportionate investigative tool when used appropriately, one that returns candidates in seconds where manual comparison used to take days or weeks, and that has reduced the need for forces to issue public appeals. The inspection also found that the public is broadly supportive. HM Inspector Kathryn Stone said the technology had transformed suspect identification, while adding that "decisions are still subject to human judgment".
The report gives a practical example of what that looks like. After a knife-point robbery in Crewe, an urgent search by Cheshire Constabulary's central team produced a suspect within ten minutes, and he was arrested shortly afterwards when British Transport Police stopped a train. Cheshire's approach, in which a single trained team handles every search and its staff complete a five-day facial recognition course, was singled out by inspectors as good practice that other forces could learn from.
Measuring what the searches contribute
Where the inspectorate did find a gap, the first was in evidence of results. According to Biometric Update's account of the report, forces were often unable to show clearly how far the searches had improved performance, largely because they lacked the management information needed to link a search to the outcome of an investigation.
Use also varies a great deal from one force to the next. One force ran 29,742 searches during the year while another ran 611, and although two forces made more than 25,000 requests each, several made fewer than a thousand. Some of that difference will reflect force size and local crime patterns, but the inspectorate's own reading is that a number of forces are not yet making full use of a capability that could help them. Without outcome data it is difficult to say which forces have the balance right, and that works in both directions: the same missing figures that prevent critics from testing the technology also prevent policing from demonstrating its value. West Yorkshire Police's case management software, FaceReveal, was highlighted as an example of searches and results being tracked together.
Audit arrangements have not kept pace
The second gap concerns audit. Searches of the national database come with audit requirements, and local reporting of the inspection states that only two forces had audited the correct number of searches and that seven had not appointed an auditor. Big Brother Watch, responding to the report, described this as forces not recording how the tool is used.
This matters mainly because human review and after-the-fact audit are the principal safeguards for this kind of searching, so the system depends on both being carried out. It also echoes a point made in this site's analysis of licence plate cameras in the United States, which is that a good audit trail protects the organisation as much as the public, since it is the only means by which a force can show that its officers are using a system properly. The inspectorate has recommended that the audit requirements themselves be reviewed, which suggests it sees the difficulty as lying partly in how the rules were designed and not only in how forces have applied them.
Training for the people who request searches
A search result is intelligence and not identification evidence, meaning that it suggests a possible match which an investigator then has to corroborate through further enquiries. Experienced investigators and specialist database operators generally understood this, but the inspection found the understanding was weaker among less experienced officers, who are also the ones most likely to request a search.
In one force, more than half of the officers in a focus group did not know that results were for intelligence purposes only. A Home Office survey of 779 database users, published alongside the report, found that roughly a third were wrong or unsure about whether a match counts as identification evidence, and that nearly two-thirds could not recall receiving any facial recognition training in the previous year. Much of the existing guidance and training is aimed at the operators who run the searches, which leaves the wider group of investigators who ask for them less well served.
The report gives two examples of what can follow. In a shoplifting investigation a man was circulated as wanted, arrested in another force area and transferred, at which point officers saw that he did not match the image of the offender, and it emerged that no further enquiries had been made before he was circulated. In another force, inspectors were told that supervisors had directed officers to make an arrest on the basis of a potential match alone. The report presents these as illustrations of a training gap and does not suggest they are typical.
Reactions differed in emphasis. Big Brother Watch said officers were treating a match as "bona fide evidence rather than a probability score", while Ben Russell, who leads on facial recognition for the National Police Chiefs' Council, said policing remained committed to ethical, proportionate and responsible use and welcomed the inspectorate's recognition of what the technology can do.
What the report does and does not say about bias
The bias findings need stating carefully, because some of the coverage went further than the report does. Testing by the National Physical Laboratory found that, at some settings, the algorithm behind the database search was more likely to return women, Asian and Black people, and people aged 21 to 29 in lists of potential matches. Biometric Update reports that the threshold was raised to reduce those false positives, that doing so also reduced the number of successful matches, and that it was lowered again in November 2024 following operational feedback. That sequence illustrates a real trade-off for the people running the system, since a setting that returns fewer false candidates also misses more genuine ones.
The inspectorate acknowledged that policing has taken steps to address the bias, including training and the requirement that every result is reviewed by a person. It also said that it could not establish with certainty whether the bias had contributed to any wrongful arrests. That leaves the question open in both directions, and it is likely to stay open until outcomes are recorded more consistently, because the information needed to answer it is the same information that is currently missing.
A replacement is on its way. The Home Office has selected a new algorithm for the national system, the National Physical Laboratory found no significant difference in its performance between demographic groups, and Cheshire and Essex have already trialled it. It is due to become available to forces next year.
Custody images held beyond their retention date
The database holds more than 20 million custody images, which are uploaded whether or not the person photographed was later charged or convicted. Inspectors found that some images which should have been deleted under the existing retention rules were still available for searches, and described retention as an area of concern.
This is a long-standing issue that predates facial recognition, and it owes a good deal to older record systems in which deleting an individual image is a manual task. What has changed is the consequence, because an image that remains in the gallery can now be returned as a candidate in an investigation, which gives the retention rules a practical importance they did not have when photographs were simply filed.
Sixteen recommendations, most of them due in 2027
The inspectorate made 16 recommendations, addressed variously to chief constables, the National Police Chiefs' Council, the College of Policing and the Home Office. As reported, they include a local policy in every force and a review of audit requirements by the middle of 2027, national requirements for management information and an independent oversight framework by the end of that year, and mandatory training for everyone involved in these searches by 31 December 2027.
These sit alongside the statutory framework the Home Office announced on 4 November, which would create a regulator and a code of practice through the Police Reform Bill. The inspectorate noted that legislation has not kept up with the technology and that this makes it harder for forces to be sure of the lawful basis for what they are doing, so some of the inconsistency it found is a product of forces working without clear national rules. In the meantime the searches will continue at a rate of roughly 800 a day while the improvements are put in place.
One of the seven inspected forces, Dorset, has since been reported locally as saying that an audit of its own use found no issues, which is a useful reminder that the national picture is an average and that practice in individual forces varies considerably.
Figures that would settle most of the questions
Much of what remains uncertain after the inspection could be resolved by a small number of figures, none of which would need new legislation to collect:
- each force's annual number of searches, alongside the number that were audited;
- a national count of matches that contributed to a charge, and of matches later shown to be wrong;
- the number of custody images deleted each year, and the number still held past their review date;
- the operating threshold in use, with the date of and reason for any change to it.
Publishing figures of that kind would serve policing at least as well as it would serve its critics. Retrospective facial recognition has been part of the Police National Database since 2013 and is among the most heavily used pieces of police technology in the country, and the inspectorate's view is that it earns its place. The work that remains is to build the records that would allow forces to show as much.