The Brief
Society & Ethics 4 min read

Facial Recognition as Evidence: The Mistake That Gets People Arrested

NAVION

Share

A formal inspection of how British police use facial recognition technology has produced a finding that cuts to the heart of how AI systems can go wrong in practice. The problem is not simply that the technology makes mistakes. It is that officers sometimes treat those mistakes as facts.

When a Probability Score Becomes a Warrant

His Majesty’s Inspectorate of Constabulary published a report examining retrospective facial recognition, the form of the technology used to search existing databases of images rather than scan crowds in real time. Even this less controversial application, the report found, carries serious risks when the humans operating it do not understand what its outputs actually mean.

The core issue is a category error. Facial recognition results are intelligence, a lead to be investigated, not evidence that can justify enforcement action on its own. Yet the report found that some investigators did not know this distinction existed. In a small number of cases, officers acted on potential matches without carrying out sufficient follow-up checks. The result, in those cases, was what the report calls “inappropriate investigative action.”

This is what most coverage of facial recognition misses. The debate tends to focus on whether the technology should exist at all. The report shifts the question: even where the technology is permitted, are the people using it equipped to use it correctly? The answer, in at least some documented cases, is no.

Bias, Bad Images, and Missing Safeguards

The report identifies several compounding problems. The algorithm itself carries bias, including a higher likelihood of wrongly flagging people from ethnic minorities as suspects. This is not a new finding in the broader field of facial recognition research, but its presence in an official inspection of active police use gives it particular weight.

The quality of the images fed into the system also matters. Auditors found cases where officers uploaded blurred images or images containing multiple faces, both of which reduce the reliability of any match the system returns. The technology can only work as well as the inputs it receives.

Jasleen Chaggar of Big Brother Watch raised a further concern: police are not consistently logging their use of retrospective facial recognition. This means individuals could become suspects in criminal investigations without any transparent record of why they were flagged, and without any practical way to challenge the process.

The report also found that custody images which should have been deleted were in fact retained, creating an additional risk to public confidence. And underpinning all of this is a legislative gap. As the report states, the law has not kept pace with the technology. There is no comprehensive statutory framework governing how facial recognition can be used, which makes it harder for police forces to understand what is legally permitted and what safeguards are required.

Professor Pete Fussey of the University of Southampton, formerly a police adviser on facial recognition, put the legal dimension plainly: meaningful human scrutiny of AI systems is already a legal requirement, yet the report documents cases where investigators accepted the computer’s output without further verification.

What This Reveals About AI in High-Stakes Environments

The facial recognition case is a specific instance of a much broader challenge. When AI systems are deployed in environments where their outputs can trigger real-world consequences, such as an arrest, a medical decision, or a financial action, the humans in the loop need to understand precisely what the system is and is not telling them.

An AI match is not a conclusion. It is a signal that warrants further inquiry. The moment that distinction collapses, the system stops augmenting human judgment and starts replacing it, and not in a way that serves anyone well. Kathryn Stone, His Majesty’s Inspector of Constabulary, made this explicit: decisions must remain subject to human judgment, and appropriate safeguards must be demonstrably in place.

The report’s call for a clearer statutory framework, stronger national standards, and improved governance is not bureaucratic caution. It reflects a genuine gap between how fast the technology has been adopted and how slowly the institutional structures around it have developed. Technology moves faster than law. That gap is where harm tends to occur.

Ben Russell of the National Police Chiefs’ Council acknowledged the report while affirming policing’s commitment to ethical and proportionate use of the technology. The tension between that stated commitment and the documented failures is exactly what the report asks forces to resolve.

In Short

Facial recognition can support criminal investigations, but only if the people using it understand that a match is a starting point, not a conclusion. The inspection found that this understanding is not universal among officers, that the algorithm carries bias against ethnic minorities, and that the legal framework governing the technology’s use remains incomplete. The risk is not just wrongful arrest. It is the erosion of public trust in a tool that, used correctly, could have genuine value.

Based on reporting from The Guardian.

Written by

NAVION