UK judge: Home Office refused asylum on AI-hallucinated doc

A senior UK judge has said a Home Office letter refusing a Moroccan woman's asylum claim looks like it was written with help from a generative AI tool, and that the authoritative country policy document the refusal rested on may never have existed. The Guardian reports that in a 14 July judgment the judge wrote the letter "bears hallmarks consistent with the use of artificial intelligence" and that references to a country information note on Morocco dated July 2021 in that letter "are a result of an AI 'hallucination'".

The underlying case is grim. The woman said she had fled Morocco with her child after a forced underage marriage and extreme violence including rape, and that her husband, a powerful previously convicted criminal, would kill her if she returned. The refusal leaned on what it described as a Country Policy Information Note, a document type judges normally treat as authoritative, to conclude Morocco would be safe for her. When lawyers asked to see it, the Home Office first sent a link to an archive where the document could not be found, then produced a different document, a piece of unpublished internal guidance known as a country information note. The judge concluded "It appears that no such CPIN has ever existed" and said relying on a document that did not exist was "analogous to relying on bogus evidence".

Strip out the immigration politics and the operational point is what should worry anyone shipping AI into a decision workflow. The problem is not that a model made something up, which is well understood by now. It is that a fabricated citation moved through drafting, review and issuance as if it were a real government document, and was then used to reject a life-or-death claim. The controls that were supposed to catch it did not.

The honest caveat is that the judgment stops at hallmarks consistent with AI use and what may transpire; it does not name a specific model, tool or caseworker workflow, and the Home Office has not conceded a tool was used in this letter. Separate reporting places the department's own generative tools, including an Asylum Case Summarisation tool and an Asylum Policy Search tool that reportedly use GPT-4o, in the asylum pipeline, but the direct link from those tools to this specific refusal is not established in the judgment itself.

What is worth watching from here is quieter than the headline. Every previous refusal that cited a similarly obscure country note is now a candidate for challenge, and legal aid firms will notice. Vendors selling grounded, citation-verified retrieval have an easier sales conversation into government than they did last week. And whichever agency moves first to require that any AI-drafted decision letter carry a verifiable source trail will set the pattern the rest end up copying.