UK Home Office asylum AI fails one in five checks as testing stops

An Open Rights Group investigation found the UK Home Office's GPT-4o-based asylum tools produce errors in 15–22% of spot-checked outputs, up from 9% in the pilot. Quality testing on the second tool has stopped, and oversight covers only 0.3% of roughly 70,000 cases processed between April and July 2026.

• 45 of 203 sampled ACS summaries (22%) failed quality review
• Second-stage review found 15% failures versus 9% in the pilot
• Sampling covered 0.3% of ~70,000 cases from April to July 2026
• Both tools run on GPT-4o and are not disclosed to applicants