Show How ChatGPT Enters an Asylum Claim

The UK Home Office uses tools built on OpenAI’s GPT-4 technology to help process asylum claims, Wired reports. The Independent Chief Inspector of Borders and Immigration is considering scrutiny of the tools. According to Open Rights Group campaigner Sara Alsherif, the inspectorate requested more information from the nonprofit while preparing next year’s inspection priorities. No investigation has been confirmed.

The tools are Asylum Case Summarisation and Asylum Policy Search. Documents obtained by Open Rights Group describe their use to summarize interview transcripts and information about applicants’ countries of origin. Wired reports that the tools use OpenAI’s 2023 GPT-4 model and Microsoft Foundry. These tasks concern the evidence and information a caseworker consults. They are not, by themselves, recommendations or decisions on a claim.

“Asylum decisions are made by trained caseworkers,” a Home Office spokesperson told Wired. The department says the tools assist with navigating interview records and policy information, cannot be used in isolation, and do not decide claims. That distinction matters. It still leaves the checking procedure unspecified. An assistant needs permissions and supervision, not just a modest job title.

A summary can influence an assessment without issuing an instruction. If a caseworker relies on a shortened interview record, the selection of details matters. A qualification, a chronology, or an explanation of an apparent contradiction may affect how the account reads. These are risks to test, not proof that the tools have distorted a particular applicant’s evidence.

There is a documented reason to examine reliability. Wired reports that Open Rights Group found a higher share of ACS summaries unfit for use than the published pilot evaluation’s 9 percent error rate. The supplied reporting does not give a replacement figure. The inspectorate also pointed to an earlier finding that one tool’s effectiveness and reliability varied. Neither finding establishes that AI independently rejected claims.

The department says it regularly monitors performance through established quality-assurance processes and has committed to publishing updated evaluations. Useful disclosure would connect those evaluations to the working procedure: whether officials must read the full transcript, how summaries link back to source passages, and how reviewers record corrections. Faster access to an applicant’s account is useful only if officials can check what the shortened version leaves out.

The Home Office can publish the workflow, the verification instructions, and the rules governing retention of original and corrected outputs. It can identify who may reject a faulty summary, who must investigate repeated errors, and how an applicant or representative can challenge a misstatement that enters the case record. Those disclosures would make human responsibility inspectable rather than merely asserted.