UK Police Pauses Biased LFR Deployment

🔑 Enhanced Key Takeaways

• •The suspension was triggered by a University of Cambridge study in Chelmsford involving 188 actors, which found the system was 'statistically significantly more likely' to correctly identify Black participants than other groups, raising concerns about disproportionate surveillance targeting.
• •An Information Commissioner’s Office (ICO) audit revealed that Essex Police had scanned approximately 2.5 million people using the technology before the pause was enacted, despite the force previously claiming only one incorrect alert had occurred.
• •A parallel National Physical Laboratory (NPL) evaluation of the Police National Database (PND) found that at lower confidence settings, the False Positive Identification Rate (FPIR) for Black women was nearly 100 times higher than for white women.
• •The National Police Chiefs' Council (NPCC) reportedly reversed a 2024 decision to raise algorithm confidence thresholds after the change caused 'investigative leads' to plummet from 56% to 14%, prioritizing volume over bias mitigation.

🛠️ Technical Deep Dive

• •Algorithm Providers: Essex Police utilized Corsight AI and Digital Barriers; the Police National Database (PND) uses Cognitec FaceVACS-DBScan and Idemia MBSS.
• •Threshold Sensitivity: The PND system showed a 0.04% FPIR for white subjects vs. 5.5% for Black subjects at lower settings; raising the threshold to mitigate this reduced match rates by 75%.
• •Operational Metrics: The Cambridge study evaluated the True Positive Identification Rate (TPIR) at a threshold of 55, finding a 50.7% correct identification rate for watchlist targets.
• •Hardware Integration: Deployments utilize mobile vans equipped with high-definition cameras and H.265 video streams for real-time processing against watchlists of up to 10,000+ individuals.

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline