Microsoft Copilot Fabricated Match Police Intelligence
West Midlands Police in the UK were asked to advise on whether Maccabi Tel Aviv supporters could safely attend an away Europa League fixture at Aston Villa in Birmingham. The intelligence report the force produced cited "previous unrest" at a match between West Ham and Maccabi Tel Aviv as part of the case for restricting the away fans - stated as settled fact. That match never took place. No such fixture existed, and the fabricated match was carried into the report as if it were a real prior incident. For weeks the force denied that AI tools were involved. Chief Constable Craig Guildford told Parliament the erroneous item had come from a web search and social-media scraping. On 12 January he wrote to the Home Affairs Select Committee to correct the record: "I [recently] became aware that the erroneous result concerning the West Ham v Maccabi Tel Aviv match arose as result of a use of Microsoft Co Pilot." He apologised to MPs for giving them erroneous evidence about the ban. The restriction went ahead regardless: Aston Villa's home fixture was played on 6 November with no away supporters present, and the decision became a national political dispute. The same risk assessment also repeated an inflated account of Maccabi fans' behaviour at an earlier Amsterdam match - 500 to 600 fans allegedly targeting Muslim communities and "throwing random members of the public" into a river, and 5,000 officers needed to police the unrest after a figure of 1,200 had previously been given. Amsterdam police said they did not recognise those claims, and in December the Dutch police inspectorate, in a letter obtained by the BBC, said it saw "no new facts or circumstances that would warrant revising these findings". Is that right?
1 Answer
Expert: Craig Guildford, Chief Constable, West Midlands Police (letter to the Home Affairs Select Committee, 12 January 2026) Copilot did not simply get a fact wrong inside a chat window. It generated a fixture that does not exist, and that fabricated item was copied into a police intelligence report and then used to justify restricting thousands of supporters from attending a match. Three failures compounded. First, the claim was never checked against a primary record. A fixture list would have settled it in seconds; instead the output travelled straight into a report used to make a decision about people. Second, when the claim was challenged, the force defended it and denied AI involvement rather than auditing its own sourcing. Leaked minutes of the safety advisory group meetings later showed police had indicated a preference for a ban despite an "absence of intelligence" - so the AI output was not the origin of the conclusion, it was the corroboration that made a pre-existing conclusion look evidenced. Third, by the time the record was corrected the decision could not be undone. The fixture had been played without away supporters, and the episode had become a national political dispute. Guildford apologised to the committee for the erroneous evidence and resigned as chief constable in January 2026; an acting chief constable has led the force since. The standard is the same one that applies to every case on this site: when an AI output is going to be used to make a decision about people, the specific claim must be traceable to a primary record - a fixture list, a court filing, a policy document - before it is acted on. An apology issued after the harm is not a correction. Source: https://arstechnica.com/ai/2026/01/deny-deny-admit-uk-police-used-copilot-ai-hallucination-when-banning-football-fans/
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?