Errors and bias in retrospective searches

According to the findings issued on Wednesday, facial recognition errors run through the technology, and the people operating it do not always grasp that extra checks are needed before its output can be trusted. The inspection looked only at retrospective facial recognition, the variant generally judged less contentious than live systems that scan crowds for suspects in real time.

The inspectors identified several specific failures:

• •Some investigators were unaware that RFR output counts as intelligence, not evidence, and that more enquiries must follow before any enforcement step.
• •The algorithm still mislabels suspects along ethnic lines, with people from ethnic minorities more likely to be flagged in error.
• •Blurred photographs and pictures showing several faces were among the unsuitable probe images staff fed into the system, weakening the matches it could deliver.
• •In a few instances, officers gave RFR results more weight than they deserved, producing inappropriate investigative action.

The inspectors warned that mistaken identification or arrest on this basis could erode public trust in the technology.

Custody images and the missing statutory framework

Photographs of people in custody that ought to have been removed from records were still being held, a practice inspectors linked to a threat to public confidence.

The report concluded that the law has lagged behind the technology, leaving no comprehensive statutory framework for how facial recognition is used. Without one, forces struggle to establish on what legal basis they may deploy RFR and which safeguards should accompany it.

Watchdog and campaign group responses

His Majesty's Inspector of Constabulary Kathryn Stone said: "Police forces must be able to show that appropriate safeguards are in place, risks relating to bias and data retention are effectively managed". She argued that public confidence depends on a clearer statutory framework, stronger national standards and better governance and oversight.

Jasleen Chaggar of Big Brother Watch called it alarming that officers treat the technology as "bona fide evidence rather than a probability score". On her account, this habit makes it significantly more likely that innocent people will face officers at their door arresting them over crimes in towns they have never visited.

She added that forces often fail to record when they run retrospective facial recognition, leaving people unable to see why they became suspects. In her view, the algorithm cannot be run with so few safeguards and at biased settings that misidentify women, Asian and black people most often, especially while searches draw on millions of custody images held unlawfully.

Police and academic reaction

Ben Russell, the National Police Chiefs' Council's lead on facial recognition, said policing "remains committed to the ethical, proportionate and responsible use of retrospective facial recognition" when investigating crime. He welcomed that HMICFRS had recognised how much promise the technology holds.

The findings arrive at a moment when police chiefs frame facial recognition as crime fighting's next big leap forward, while sceptics call it the next huge threat to civil liberties.

Prof Pete Fussey of the University of Southampton, once an adviser to police on facial recognition, said: "Meaningful human scrutiny of AI systems is a legal requirement". He pointed to cases in the report where investigators accepted what the computer suggested without carrying out further checks.

Stone's view is that public confidence in the technology rests on a clearer statutory framework, tougher national standards and stronger governance and oversight.