REPORTED: These figures come from the Open Rights Group (ORG), a digital rights campaign group, based on its Freedom of Information requests and replies from ministers. The Home Office has not published them itself.
ORG says 45 of 203 AI-written summaries of asylum interviews failed a first check by Home Office experts. That is 22%. In the Home Office’s own pilot, 9% of summaries were judged not fit for use [1][2].
ORG published its briefing on 12 October 2026. It is calling for both tools to be suspended and audited independently [1][2].
What are the tools? (REPORTED)
ORG says the Home Office uses two AI tools in asylum cases, “both built on a large language model (GPT-4o via Azure AI Foundry)” [1]:
- Asylum Case Summarisation (ACS) summarises asylum interview transcripts for the officials who decide claims.
- Asylum Policy Search (APS) is a chat-style tool for searching the Home Office’s information on the countries people are fleeing.
According to ORG, APS “was fully rolled out to all asylum decision-makers from July 2025” and ACS “was rolled out to all asylum caseworkers from April 2026” [1].
A large language model writes new text, so it is not just copying from the record. ORG says this means the tools “can filter, reframe or omit facts relevant to a decision” [1]. In an asylum interview, one missing detail can change the outcome.
The checks (REPORTED)
After ACS went live, Home Office subject matter experts (SMEs) spot-checked a sample of summaries. ORG says [1]:
“In a first-stage SME check on 203 summaries, 139 passed, 19 were flagged for further review, and 45 failed (22%). Of 56 second-stage reviews, 30 failed.”
ORG concludes that “across both stage reviews 22% and 15% of sampled summaries were assessed as not fit for use – significantly higher than the 9% figure in the Home Office’s published pilot evaluation” [1]. The 15% matches the 30 second-stage failures as a share of all 203 sampled summaries.
The sample was small. ORG says “only 203 were checked by Subject Matter Experts during this period – just 0.3% of generated text for the cases” between April and July [1].
The pilot figure (CONFIRMED document; figure as cited by ORG)
The Home Office published “Evaluation of AI trials in the asylum decision making process” on GOV.UK on 29 April 2025 [3]. ORG says that evaluation “found 9% of the summaries it produced were inaccurate or had missing information and had to be removed from the pilot”. It also found 23% of caseworkers were not fully confident in ACS output, and 5% were not confident in APS accuracy [1].
Slower decisions (REPORTED; no cause established)
ORG says its FOI requests showed that “since April 2026, the average monthly number of initial substantive decisions has decreased by about 42% compared with the same period in 2025”. Over the same period, the average (mean) time from interview to decision rose by 29% [1].
These are ORG’s figures. They do not show that the AI tools caused the slowdown. ORG uses them to question the Home Office’s claims that the tools save time [1].
What ORG wants (REPORTED)
ORG is asking the Home Office to [1]:
- Suspend both AI tools in asylum casework and take steps to help asylum applicants whose cases were affected.
- Commission an independent evaluation and audit of how accurate the tools are in live use, and publish the results.
- Ask the Independent Chief Inspector of Borders and Immigration to inspect the tools.
The Home Office’s response
The Independent reported the same check figures on 8 October 2026. It quoted a Home Office spokesperson [4]: “Asylum decisions are made by trained decision-makers, not artificial intelligence. Our AI tools support caseworkers by helping them navigate interview records and policy information, but they cannot be used in isolation and do not make decisions on claims.” The spokesperson added: “We regularly monitor the performance of these tools alongside our established quality-assurance processes.”
The Register covered ORG’s report on 13 October 2026 and said “The Home Office did not respond to The Register’s questions” [5].
What this means
- Spot-checks found more errors than the pilot did. By ORG’s account, more than one in five of the summaries checked failed the first review.
- Very few were checked. A 0.3% sample cannot show how accurate the tool is across all cases, whether that turns out better or worse.
- The Home Office says humans decide. Its position is that officials, not the AI, decide claims. ORG argues that a wrong summary can still shape a human’s decision.
- This is a campaign group’s analysis. The Home Office has not published these figures or an independent audit.
Sources
- Open Rights Group, “The Use of ChatGPT in Asylum Decisions: Errors, inaccuracies & Home Office incompetence”, briefing, 12 October 2026. https://www.openrightsgroup.org/publications/the-use-of-chatgpt-in-asylum-decisions/ (PDF: https://www.openrightsgroup.org/app/uploads/2026/10/ChatGPT_in_Asylum_Decisions_Errors_inaccuracies_and_incompetence.pdf)
- Open Rights Group, press release, 12 October 2026. https://www.openrightsgroup.org/press-releases/org-investigation-into-use-of-chatgpt-in-asylum-decisions-highlights-errors-inaccuracies-and-home-office-incompetence/
- Home Office, “Evaluation of AI trials in the asylum decision making process”, GOV.UK, 29 April 2025. https://www.gov.uk/government/publications/evaluation-of-ai-trials-in-the-asylum-decision-making-process
- The Independent (Holly Bancroft), “AI tool used in thousands of asylum claims failing basic accuracy tests”, 8 October 2026. https://www.independent.co.uk/news/uk/home-news/ai-asylum-claims-backlog-home-office-artificial-intelligence-b3055219.html
- The Register (Carly Page), “Home Office asylum AI flunks one in five accuracy checks”, 13 October 2026. https://www.theregister.com/public-sector/2026/10/13/home-office-asylum-ai-flunks-one-in-five-accuracy-checks/5302649
