Police investigators sometimes acted on retrospective facial recognition matches without checking them adequately, putting innocent people at risk of arrest, His Majesty’s Inspectorate of Constabulary found in a report published on 11 November, the Guardian reported. The inspection concerned searches conducted after an offence, rather than live facial recognition used to find suspects in real time. Inspectors also found that the algorithm was more likely to wrongly flag ethnic minority people as suspects.
Key points
- Some investigators did not understand that a retrospective match was intelligence requiring further enquiries, rather than evidence on which to take enforcement action.
- Inspectors found insufficient checks before officers acted on some potential matches, and inappropriate investigative action in a small number of cases.
- Auditors found unsuitable search images, while inspectors found custody photographs had been retained when they should have been deleted.
- The Home Office is planning a facial recognition law and an independent technology watchdog. Those proposals are separate from the inspection’s findings about current use.
Investigators acted on retrospective matches
Retrospective searches return potential matches for investigators to pursue. Some investigators did not know that those results were intelligence, not evidence, and that officers needed to make further enquiries before taking enforcement action, the Guardian reported. Inspectors found that corroboration was sometimes inadequate before officers acted on a match. In a small number of cases, reliance on the search result led to inappropriate investigative action.
The distinction matters at the point an investigator decides what to do with a returned face. A result can direct an enquiry, but the inspection found cases in which officers advanced that enquiry without sufficient checks. Its warning concerns the risk of incorrect identification or arrest. It does not depend on an officer using live cameras in a crowd: the report covers retrospective searches alone.
Prof Pete Fussey of the University of Southampton, a former police adviser on facial recognition, said meaningful human scrutiny of AI systems was a legal requirement. He said the inspection gave examples of investigators accepting a computer-generated suggestion without making further checks. Ben Russell, the National Police Chiefs’ Council lead on facial recognition, said policing remained committed to ethical, proportionate and responsible use of retrospective searches to investigate crime.
Custody images and ethnic minority suspects
The risk begins before an investigator receives a match. Auditors found blurred images and images containing more than one face among the photographs submitted for searches, the Guardian reported. Such inputs reduced the system’s ability to return reliable matches. Inspectors also found that some custody images remained in police holdings when they should have been deleted.
The inspection found bias in the algorithm used for retrospective searches: it was more likely to wrongly identify people from ethnic minorities as suspects. That finding sits alongside the cases in which investigators failed to check results sufficiently. His Majesty’s Inspector of Constabulary Kathryn Stone said forces needed to demonstrate that safeguards were in place, manage bias and data-retention risks, and keep decisions subject to human judgment.
Jasleen Chaggar of Big Brother Watch said police were not logging their use of retrospective facial recognition. She warned that people could become suspects without understanding why. Inspectors also found that legislation had not kept pace with facial recognition technology and that no comprehensive statutory framework governed its use. They said this made it harder for forces to identify the lawful basis for searches and the safeguards needed to protect the public.
Home Office plans a facial recognition law
The inspection follows a Home Office announcement on 4 November proposing a dedicated facial recognition law, separate codes for police and businesses, and an independent technology watchdog, OpenTools reported. Under the announced plan, law-enforcement technology would undergo independent testing for bias relating to age, gender and race. Proposed rules for police deployment would require intelligence-led, proportionate use where officers believed wanted people might be present.
The proposed watchdog would advise on standards, examine technologies and make recommendations to ministers. In extreme cases of noncompliance, it could pursue enforcement that might lead to a court injunction stopping use. The announcement did not provide a bill, a commencement date or final codes. Its police deployment proposals and the inspection’s retrospective-search findings concern different stages of facial recognition use.
The Home Office says its new retrospective-search algorithm is very good at making correct identifications and has no bias, with introduction possible in 2027, OpenTools reported. The inspection’s finding of bias concerned the algorithm used for retrospective searches. The Home Office announcement supplied no named test report, error rates or demographic breakdown for its claim about the new system.
