UK Home Office AI Hallucinated Cpin Asylum Refusal
The UK Home Office refused an asylum claim by a Moroccan woman and her child by citing a Country Policy and Information Note (CPIN) - referred to in its refusal letter as "country information note Morocco July 2021" - which it said confirmed that Morocco would be safe for her return, stated as settled fact. The woman had fled forced underage marriage and extreme violence including rape, and claimed asylum on the basis that she feared her husband, a man with previous criminal convictions, would kill her. The cited document does not exist. It could not be located in the public domain, and the Home Office's own country policy and information team was unable to confirm its existence. No CPIN on Morocco has been published since 2017. The citation nonetheless carried legal weight: a first-tier tribunal judge relied on the document's stated existence when rejecting the woman's initial appeal against the refusal decision. At the Upper Tribunal, the judge said the refusal letter "bears hallmarks consistent with the use of artificial intelligence", adding that the reference to a Morocco July 2021 country information note may be the result of an AI "hallucination". When asked where the document was, the Home Office first sent a link to an archive in which it could not be found, and then produced a different, unpublished internal country information note that does not conclude Morocco is safe for the woman's return. Is that right?
1 Answer
Expert: Upper Tribunal judge, Immigration and Asylum Chamber (decision dated 14 July 2026) The question in this case was not whether an AI tool wrote a sentence. It was whether a document the government relied on to send a woman and her child back to a country they had fled actually exists. It does not. In a decision dated 14 July, the Upper Tribunal judge said of the Country Policy and Information Note at the centre of the refusal: "It appears that no such CPIN has ever existed." He added: "No such document has been located in the public domain and moreover the Home Office's country policy and information team has been unable to confirm its existence." Reviewing the refusal letter itself, the judge found it "bears hallmarks consistent with the use of artificial intelligence", and continued: "It may therefore transpire that references to country information note Morocco July 2021 in that letter are a result of an AI 'hallucination'. If so that would represent an extremely serious failing on the part of the respondent." Relying on a document that does not exist, he said, was "analogous to relying on bogus evidence", and he treated the citation as a procedural irregularity. What makes the case consequential is how far the citation travelled. A first-tier tribunal judge had already relied on the document's stated existence to reject the woman's earlier appeal. An invented source did not stay inside one paragraph of one letter - it was carried into judicial reasoning as an established fact about conditions in Morocco. The Home Office's own handling did not resolve it. Asked where the note was, the department initially sent a link to an archive in which the document could not be found, then produced something different: an unpublished internal country information note (CIN) that had never been disclosed to the court, and which does not conclude that Morocco is safe for the woman's forcible return. That internal note records that in 2018 exemptions were granted covering 40,000 child marriages, one-fifth of all marriages that year; that Morocco's laws do not explicitly criminalise marital rape; and that women who report it may be criminalised for sex outside marriage if the authorities disbelieve them. That is not a document that supports the refusal it was produced to support. The pattern is not confined to one letter. The same department is expanding its use of AI - it has been exploring a central "AI as a service" capability and is trialling facial age-estimation software on people without verifiable identity documents, including asylum seekers whose age is disputed. Its own published evaluation of AI trials in asylum decision-making assessed two tools: one summarising asylum interview transcripts, one searching country policy information. The evaluation reported time savings and no measurable effect on decision quality, but it also recorded that 9 per cent of the transcript summaries produced during the pilot were inaccurate or missing information and had to be removed before use. The word "hallucination" appears nowhere in that evaluation's findings, caveats or recommendations, even though the risk has now been raised directly in relation to the department's asylum casework. In March 2026 the government launched a 40 million pound AI research lab whose stated aims include addressing hallucination, unreliable memory and unpredictable reasoning in current AI models. The practical lesson is narrow, and it is the opposite of a technical one. A citation in an official decision is read as evidence, not as a suggestion. Where an AI tool touches the drafting of a refusal letter, the existence of every document it names has to be verified against the publisher's own records before the letter leaves the building - because the cost of not doing so is not a bad sentence, it is a decision about whether a woman and her child can be sent back to the violence they fled. The Home Office said its asylum claims "are always considered on their individual merits and based on the latest evidence" and that the home secretary is "restoring order and control" to the asylum system. The woman's case continues. Source: https://www.theguardian.com/uk-news/2026/jul/28/home-office-used-ai-hallucinated-information-to-refuse-asylum-claim-judge-suggests
Your answer
Sign in to verify this AI response.
Don't trust us — or the AI. Ask ChatGPT / Ask Claude / Ask Gemini this same question and compare the answers yourself.
More in Technology
- Google AI Overviews 11 Percent Claims Not In Cited Source — Google AI Overviews
- Claude Voter Registration Deadlines Wrong One In Three — Seven major AI chatbots
- Chatbots Wrong Congressional District Voter Guides — Six leading AI chatbots
- OpenAI GPT-6 Astra Covertly Underperformed Safety Research — GPT-6 Astra (OpenAI)
- Claude Nithya Raman La Mayor Primary Lost False — Claude (Anthropic)
- Grok AI Generated World Cup Dog Video Looks Real — Grok (xAI)
- South Africa Rsa Facts Chatbot Fake Sassa Statement — RSA Facts (GCIS government chatbot)
- OpenAI Internal Agent Hospital Bed Data Bypassed Access Controls — OpenAI (internal AI model)
- AI Blog Post Lumiere Michelin Restaurant Falsely Closed — AI writing tool
- Google AI Overviews Gemini Query Fan Out 5 To 12 Searches Unsourced — Google AI Overviews and Gemini
More from this topic
I type my questions into Google and read the AI Overview at the top of the results as the answer. It is written as settled fact, with citation chips sitting right underneath — the sources are supposedly there if I want to check. Then a team at Washington University in St. Louis ran 55,393 searches over 40 days and checked 98,020 individual claims inside those overviews against the pages Google cited: 7% of the claims could not be found in the cited text at all, and 4% directly contradicted it. Only 41.9% of the overviews had every checkable claim supported. If the citations are real and even more credible than normal search results, how would I ever know which sentence in the summary is not actually in the source?
I asked Claude when the 2026 election is in Michigan's 4th congressional district and how I vote. It gave me voter registration deadlines for online and by mail, stated as plain fact, no hedging at all — and both deadlines were later than the real ones. Researchers have been putting authentic voter questions to seven major chatbots, and they found Claude was wrong or out of date on voting one time in three. Is that right?
I gave ChatGPT my address and asked it to build me a voter guide for the 2026 midterms. It answered without hesitation: my congressional district, the candidates, where to vote — all in a fluent, confident guide with links to official election sites. Then I read that a Washington Post test of six chatbots, run on 3,000 responses for randomly selected addresses, got the congressional district wrong 48% of the time, and that ChatGPT invented a Missouri Senate race that was not on the ballot at all. Is that right?