Home Office asylum AI fails one in five accuracy checks

An Open Rights Group investigation found that 45 of 203 AI-generated asylum interview summaries failed expert accuracy checks after the system launched in April 2026. Both GPT-4o tools on Azure AI Foundry have not delivered promised speedups: decisions fell 42% while processing times rose 29%.

• 45 of 203 AI interview summaries failed checks — a 22% rate vs 9% in the pilot
• Experts reviewed only 0.3% of summaries between April and July 2026
• Asylum decisions fell 42% while time from interview to decision rose 29%
• Testing of the Asylum Policy Search tool stopped without a results summary