2.5 Million Faces Scanned. Wrong Settings.

Essex Police borrowed accuracy thresholds from a different algorithm and called it good enough. It wasn't.

Security surveillance cameras mounted on a wall

TL;DR: Essex Police scanned 2.5 million faces across the county using live facial recognition before pausing in late 2025. The problem: they used a confidence threshold borrowed from NEC's Neoface algorithm, not the Corsight Apollo 4 system they actually deployed. Civil society groups caught the error. Now the Information Commissioner's Office and Cambridge University are auditing the mess. The force admits to "potential accuracy and bias risks" but hasn't said when (or if) they'll restart.

They Didn't Test the Algorithm They Used

Here's the short version: Essex Police deployed Corsight's Apollo 4 facial recognition system across the county. Before pointing cameras at millions of people, they needed to set a "confidence threshold": the cutoff point where the algorithm says "that's a match."

Set it too low, and you get false positives. Innocent people flagged as criminals. Set it too high, and the system misses actual matches.

Essex Police used a 0.6 threshold. There's just one problem: that number came from NEC's Neoface system, a completely different algorithm used by other UK police forces.[1]

They applied settings from one system to another and didn't bother to check if they worked.

"The 0.6 confidence threshold was derived from NEC's Neoface algorithm, employed by other UK forces, rather than from Corsight's Apollo 4 system that Essex Police actually deployed," Big Brother Watch noted after reviewing the force's methodology.[2]

2.5 Million Faces Before Anyone Noticed

Essex Police ran live facial recognition deployments throughout 2025. The last recorded deployment was August 26. By the time they paused, cameras had scanned 2.5 million people across the county.[3]

That's 2.5 million faces run through an algorithm with untested accuracy settings.

Jake Hurfurt, Head of Research and Investigations at Big Brother Watch, put it bluntly: "Essex Police accepted the facial recognition company's claims without independent verification of the algorithm's accuracy."[3]

The force didn't test before deployment. They just assumed it worked.

What the Audits Found

Multiple independent evaluations landed in March 2026, painting a complicated picture.

National Physical Laboratory

The NPL tested Corsight Apollo 4 in controlled conditions:[2]

  • True Positive Rate: 89% (identified correct matches)
  • False Positive Rate: 0.017% with an 18,000-image watchlist (1 in 5,700)
  • Demographic analysis: No "statistically significant" bias at the 0.05 level

On paper, those numbers look okay. But lab conditions aren't the street.

Cambridge University (Real-World Deployment)

Cambridge evaluated how the system actually performed in Essex Police operations. At the operational threshold of 55:[2]

  • True match rate: 50.7%, the system correctly identified roughly half of people on watchlists
  • Incorrect identifications: "Extremely rare"
  • Demographic findings: System was "more likely to correctly identify men and Black participants than women and participants from other ethnic groups"

Read that again: the system works better on some demographics than others. That's bias, even if the overall error rate is low.

ICO Audit

The Information Commissioner's Office conducted a data protection audit as part of its "rolling series of audits of police forces deploying live facial recognition."[4] The ICO isn't pulling punches: "If forces can't get governance right, it's unlikely they'll get FRT right."

The Bias Problem Runs Deeper

Essex isn't alone. In December 2025, the Home Office identified "historic bias" in the algorithm used for retrospective facial recognition searches on the Police National Database.[4]

ICO research shows the public knows this matters: 53% prioritize accuracy and 33% want safeguards against bias in facial recognition systems.[4]

The ICO's own analysis identifies three sources of facial recognition bias:

Technology Design

How the algorithm was built and what data it was trained on.

Training Data

If training images skew toward certain demographics, accuracy follows.

Watchlist Composition

Who's on the list determines who gets flagged. Biased policing creates biased watchlists.

Did It Even Reduce Crime?

Here's the part Essex Police won't highlight: Cambridge University's evaluation found the technology had "no noticeable crime reduction impact during or after deployment."[3]

2.5 million faces scanned. Untested accuracy. Demographic bias. And no measurable effect on crime.

Big Brother Watch's assessment: "authoritarian, inaccurate, and ineffective."[3]

What Happens Now

Essex Police has paused live facial recognition deployments. There are currently no upcoming deployments planned.[5]

But "paused" isn't "stopped." The evaluations from NPL and Cambridge could pave the way for a relaunch once Essex Police addresses the threshold calibration problem and governance concerns.

Meanwhile, the government plans to expand live facial recognition fivefold across the UK. Big Brother Watch is calling for a complete halt: "LFR as a tool of general mass surveillance has no place in a democracy like Britain."[3]

What You Can Do

Know Your Rights

In the UK, you have no legal obligation to look at facial recognition cameras. You can cover your face, though police may find that suspicious.

Check Deployment Schedules

Some forces publish planned deployment locations. Essex Police lists them on their official page.

Support Digital Rights Groups

Big Brother Watch and Liberty are fighting facial recognition expansion through courts and parliament.

File Subject Access Requests

You can request what data police hold on you. If you were scanned, you have a right to know.

The Bottom Line

Essex Police deployed facial recognition cameras across their county, scanning 2.5 million faces. They used accuracy settings from a completely different system. They didn't test whether those settings worked with their actual algorithm.

Civil rights groups caught the mistake. Multiple audits are now underway. The force has paused deployments.

But the cameras are still there. The software is still installed. And the government wants to expand this fivefold.

When the police scan your face with untested technology, they're not protecting you. They're experimenting on you.

References

  1. ID Tech Wire: Essex Police Paused Facial Recognition Over Bias Risk as ICO Audit and Independent Evaluations Land (March 2026)
  2. Biometric Update: Evaluations of Corsight live facial recognition follow Essex Police trial (March 2026)
  3. Big Brother Watch: Comments on Essex Police Pausing Live Facial Recognition (March 2026)
  4. Information Commissioner's Office: Why data protection lies at the heart of responsible police use of facial recognition technology (March 2026)
  5. Essex Police: Live Facial Recognition Official Page