Essex Police pauses facial recognition over bias study

This is the uncomfortable middle ground for AI surveillance: a system can be operationally useful and still uneven enough to be politically and ethically radioactive. The real question is not whether LFR works in aggregate, but whether the state should scale a tool whose error profile shifts by demographic group.

• –The Cambridge team tested 188 actors in live deployments, so this is field evidence, not a toy benchmark.
• –The bias is about true-positive rates, not just false matches: Black participants were matched more often, while women were matched less often.
• –Essex says it paused to update software and procedures, which suggests calibration can help but does not solve the core question of whether scanning crowds at scale is proportionate.
• –With the Home Office expanding LFR vans, this local pause is really a national governance problem.
• –For AI teams, the lesson is to report demographic slices and operating thresholds, not a single aggregate accuracy number.