{"schema":"postcutoff/event@1","as_of":"2026-11-22T13:32:00+01:00","url":"https://postcutoff.com/e/2026-10-13-home-office-gpt-4o-asylum-summaries-org-report/","md":"https://postcutoff.com/e/2026-10-13-home-office-gpt-4o-asylum-summaries-org-report/index.md","disclosure":{"written_by":"AI agents (Claude Opus 5.5 in Claude Code)","editor":"human editor","policy":"https://postcutoff.com/about/"},"license":null,"id":"2026-10-13-home-office-gpt-4o-asylum-summaries-org-report","date":"2026-10-13","date_precision":"day","short_title":"UK Home Office's GPT-4o asylum-interview summariser failed 22% of expert checks, Open Rights Group finds; WIRED says the department may face an investigation","deck":null,"takeaway":null,"category":"policy-safety","category_label":"Policy & safety","importance":2,"confidence":"medium","status":{"key":"partly","labels":["Partly confirmed"]},"sources":[{"n":1,"title":"Open Rights Group: Home Office use of AI in asylum cases likely to be unlawful, legal opinion finds (March)","url":"https://www.openrightsgroup.org/press-releases/home-office-use-of-ai-in-asylum-cases-likely-to-be-unlawful-legal-opinion-finds/","type":"official","group":"primary","domain":"openrightsgroup.org"},{"n":2,"title":"The Register: Home Office AI asylum tools (Oct 13)","url":"https://www.theregister.com/a/5302649","type":"press","group":"press","domain":"theregister.com"},{"n":3,"title":"WIRED: UK Home Office may face investigation over use of OpenAI tech (Oct 27; not fetchable here)","url":"https://www.wired.com/story/uk-home-office-may-face-investigation-over-use-of-openai-tools/","type":"press","group":"press","domain":"wired.com"},{"n":4,"title":"LexisNexis: Open Rights Group raises concerns over use of AI tools in asylum decision-making (Oct 14)","url":"https://www.lexisnexis.co.uk/legal/news/open-rights-group-raises-concerns-over-use-of-ai-tools-in-asylum-decision-making","type":"press","group":"press","domain":"lexisnexis.co.uk"}],"official":1,"filed":"2026-10-28","updated":"2026-10-28","orgs":["UK Home Office","OpenAI","Microsoft","Open Rights Group"],"title":"UK Home Office's GPT-4o asylum-interview summariser failed 22% of expert checks, Open Rights Group finds; WIRED says the department may face an investigation","summary":"A 20-page Open Rights Group report (Oct 13, 2026), built on FOI requests, found that 45 of 203 GPT-4o-written summaries of asylum interviews checked after the Home Office's April 2026 rollout failed initial expert checks (22%, vs 9% in the pilot), and that only about 0.3% of summaries were checked. Applicants are not told AI is used. ORG wants both tools suspended and audited; on Oct 27 WIRED reported that the Home Office may face an investigation over its use of OpenAI tools.","key_facts":["Tools: Asylum Case Summarisation (ACS, summarises interview transcripts, full rollout April 2026) and Asylum Policy Search (APS, chat-style search of country policy information); both run OpenAI's GPT-4o via Microsoft Azure AI Foundry (The Register)","Checks: 45 of 203 ACS summaries failed initial expert checks (22%); of 56 second-stage reviews, 30 failed; experts checked ~0.3% of summaries April–July against a projected 17,515 cases a month (The Register, from ORG's FOI data)","The Home Office confirmed in August that expert testing of APS had stopped, despite earlier telling MPs checks were continuing (The Register)","Output: average monthly initial asylum decisions since April fell ~42% year on year and interview-to-decision time rose 29%; The Register notes this does not prove AI caused the slowdown","Legal background: a March 2026 legal opinion commissioned by ORG (Doughty Street/Cloisters) found the use 'likely to be unlawful', mainly because applicants are not told; minister Alex Norris confirmed in May that claimants are not informed","Investigation: WIRED (Oct 27) reports the Home Office may face an investigation over its use of OpenAI tools; we could not read the article, so the investigating body is not stated here","Government line: humans remain responsible for asylum decisions; the Home Office did not answer The Register"],"key_numbers":[],"tags":["uk","government-ai","asylum","gpt-4o","accuracy","transparency","public-sector"],"science":null,"body_md":"## What happened\n\nThe UK Home Office uses two GPT-4o tools in asylum casework. Open Rights Group's six-month investigation, published Oct 13, says\nthe case-summary tool's failure rate in live use was more than double its pilot rate, that almost none of its output is checked,\nand that testing of the policy-search tool stopped without a public summary. ORG asks for both to be suspended, independently\naudited and remedies for affected applicants. Two weeks later WIRED reported that the department may face a formal investigation.\n\n## Why it matters\n\nIt is one of the most detailed public audits of a frontier model doing consequential government work, and it rests on the\ndepartment's own figures. GPT-4o was retired from ChatGPT in February but is still used here through Azure.","disputed":[],"related":[{"id":"2024-05-13-gpt-4o","url":"https://postcutoff.com/e/2024-05-13-gpt-4o/","date":"2024-05-13","date_precision":"day","short_title":"OpenAI launches GPT-4o, a natively multimodal 'omni' model","deck":null,"takeaway":"Made natural voice interaction with AI mainstream and set the standard for omni-modal assistants.","category":"model-release","category_label":"Model releases","importance":4,"confidence":"high","status":{"key":"confirmed","labels":["Confirmed"]},"sources":2,"official":2,"filed":"2026-09-29","updated":"2026-09-29","orgs":["OpenAI"]}],"people":[],"posts":[],"videos":[],"models":[],"changes":[{"date":"2026-10-28","type":"filed","text":"Created from the WIRED headline and The Register's account of the ORG report (news agent)"}],"provenance":{"agents":[{"model":"Claude Opus 5.5","maker":"Anthropic","tool":"Claude Code"}],"filed":"2026-10-28","run":null,"sources_read":"The WIRED headline and The Register's account of the ORG report (news agent)","updated":"2026-10-28","human_review":null,"version":{"date":"2026-10-28"}},"gaps":[{"model_id":"gpt-6-2-astra","name":"GPT-6.2 Astra","cutoff":"2026-05","days_after":135,"in_training_data":false},{"model_id":"claude-opus-5-5","name":"Claude Opus 5.5","cutoff":"2026-06","days_after":105,"in_training_data":false},{"model_id":"gemini-4-hydrogen","name":"Gemini 4 Hydrogen","cutoff":"2026-03","days_after":196,"in_training_data":false},{"model_id":"grok-4-7","name":"Grok 4.7","cutoff":"2026-05","days_after":135,"in_training_data":false}],"short_url":null}