Police Watchdog Warns Facial Recognition Errors Risk Wrongful Arrests of Ethnic Minorities

A report published on Wednesday by His Majesty's Inspectorate of Constabulary warns that police officers in England and Wales risk arresting innocent people if they treat inaccurate facial recognition output as evidence. The people most exposed, it says, are those from ethnic minorities.

The inspection covers only retrospective facial recognition, known as RFR. In this use of the technology, officers run an image of an unknown person, often from a crime scene or footage, against a database to look for potential matches. It is generally seen as less controversial and less difficult to get right than live facial recognition, which scans faces in real time to find suspects.

The central finding is about how officers understand what the system gives them. According to the report, some investigators did not know that RFR results are intelligence rather than evidence, and that further enquiries are needed before enforcement action. Inspectors found that some did not always carry out enough corroborating checks before acting on potential matches. In a small number of cases, officers relied too heavily on the results, which led to inappropriate investigative action. The report says this creates a risk that people could be wrongly identified or arrested, and that this could damage public confidence in the technology.

The inspectors also found problems with the algorithm itself. Bias persists, including a greater likelihood of wrongly flagging people from ethnic minorities as suspects. Auditors found cases where users uploaded poor-quality or unsuitable probe images, including blurred pictures and images containing several faces, which reduced the system's ability to return reliable matches. The report says more needs to be done to reduce the risk of wrongful arrest.

A separate concern is data retention. Inspectors found that custody images which should have been deleted had in fact been kept, another threat to public confidence. They also point to a gap in the law: legislation has not kept pace with the technology, and there is no comprehensive statutory framework governing its use. That makes it harder for forces to understand the lawful basis for RFR and which safeguards are needed.

Kathryn Stone, His Majesty's Inspector of Constabulary, said continued public support depends on showing the technology is used properly. Forces must be able to show that safeguards are in place, that risks relating to bias and data retention are managed, and that decisions remain subject to human judgment. She called a clearer statutory framework, stronger national standards and improved governance and oversight essential.

Campaigners called the findings alarming. Jasleen Chaggar of Big Brother Watch said treating facial recognition as bona fide evidence rather than a probability score raises the risk of innocent people having officers come to arrest them for crimes in towns they have never visited, as has happened before. She added that police are not logging their use of retrospective facial recognition, so people could become suspects without knowing why. She said the algorithm should not keep being used with so few safeguards, at settings more likely to misidentify women, Asian and black people, against millions of unlawfully held custody images.

Prof Pete Fussey of the University of Southampton, formerly a police adviser on facial recognition, said meaningful human scrutiny of AI systems is a legal requirement, yet the report gives examples of investigators accepting the computer's suggestions at face value.

Ben Russell, facial recognition lead for the National Police Chiefs' Council, said policing remains committed to ethical, proportionate and responsible use of the technology to investigate crime, bring offenders to justice and keep communities safe. He said the council was pleased the inspectorate had recognised its exciting possibilities.

The argument is now over whether the inspectorate's call for a statutory framework and national standards is acted on, and whether forces begin logging how they use the technology.