Home Office asylum AI flunks one in five accuracy checks

The accuracy of AI-generated asylum interview summaries is now in question, potentially impacting individuals' cases.

A new report by the Open Rights Group (ORG) reveals that the UK Home Office's AI systems used in asylum processing have a significant accuracy problem. Over one in five AI-generated interview summaries failed initial accuracy checks, a rate that increased from the pilot phase. Furthermore, the Asylum Policy Search tool, which helps officials find country policy information, also has questionable oversight, with expert testing reportedly halted. Despite the government's assurances that humans remain in charge of decisions, ORG questions the reliability of caseworkers spotting AI errors. The report also notes a potential slowdown in asylum decisions and an increase in processing times since the AI tools were implemented, contradicting promised efficiency gains. Asylum applicants are not informed when AI is used in their cases and cannot review or correct AI-generated summaries.

The Home Office's reliance on OpenAI's GPT-4o model through Microsoft Azure AI Foundry for asylum case summarization and policy search raises concerns about the robustness of these systems. The documented failure rate of 22% in interview summaries, compared to 9% in the pilot, suggests a regression in performance or inadequate validation before full deployment. This situation highlights a broader challenge in the public sector: integrating powerful AI tools while ensuring they meet rigorous accuracy and fairness standards for sensitive decision-making processes.

The ORG's findings suggest that the promised efficiency gains from AI in asylum processing may not be materializing, with observed decreases in decision-making output and increases in processing time. This paradox, coupled with a lack of transparency for asylum seekers regarding AI usage and no avenue for recourse on AI-generated content, points to a critical need for independent auditing and stronger regulatory oversight of government AI deployments. The situation prompts a re-evaluation of how AI is validated and implemented, especially when human lives and rights are at stake.