Asylum AI summaries fail ORG checks
Open Rights Group found 22% of 203 Home Office asylum AI summaries failed a first-stage check, and 30 of 56 then rated sub-standard. ORG calls the sample non-representative; the Home Office says 0.3% of summaries were expert-checked.
The Home Office's Asylum Case Summarisation tool — ACS — got a public grade on Oct. 13, 2026, and it wasn't a good one. The Open Rights Group, which spent six months investigating two generative AI tools used in UK asylum decision-making, reported that 45 of 203 summaries it reviewed failed a first-stage check, a 22% failure share. A second-stage review then rated 30 of 56 summaries sub-standard, 54%. Those are the numbers from ORG's Oct. 12 and Oct. 13 reports. The Home Office's own pilot had already told a friendlier story: 9% of ACS summaries found inaccurate or incomplete.
What 22% and 54% actually mean
Neither figure measures "wrong" in the sense of falsified facts. ORG's failure criteria include summaries that were simply not useful, scored below 2 on relevance, or contained factual inconsistencies. A flagged summary is not automatically a false statement that misled a decision. It's a summary a reviewer judged unfit for the purpose it was generated for.
The scale is a limit you can't skip past. The 203 first-stage summaries are part of 259 checks performed by subject-matter experts from April through the end of July 2026 — 19 primary reviews were flagged for further review, and the 56 secondary reviews came out of that. ORG describes the sample as non-representative, and its reporting notes the 54% doesn't describe the whole system. The Home Office told the investigation that 0.3% of summaries in that April–July window got expert checking, so the review examined a sliver of the roughly 7,000 AI summaries generated per month on average.
The tool, the boundary, and the blind spot
ACS uses OpenAI's GPT-4 to condense asylum interview transcripts for caseworkers. It isn't deciding claims; the government says the tools support staff rather than replace human decision-makers. But a summary becomes the working record of an interview, and once a detail is dropped or blurred, the caseworker has no obvious route back to the applicant's own account. ORG found no Q&A mechanism attached to the summaries, and said the Home Office refused to hand over details on staff training, prompts, and quality-assurance testing used for the tools.
Then there's the 2025 pilot: 334 test cases against 95 comparison cases, efficiency gains observed, no impact on decision quality observed. The gaps between pilot and live deployment aren't just accuracy on paper. The pilot ran a controlled batch. Production runs caseworkers under time pressure, on real transcripts, and an internal impact assessment from October 2025 predicted an average of 17,515 asylum cases per month — a volume the checked sample can't characterize.
What's confirmed and what isn't
Minister Alex Norris has confirmed ACS remains operational, with evaluation calling for continuous monitoring. ORG wants an independent audit and a live-deployment study; no independent audit is planned, and the tools haven't been added to the Algorithmic Transparency Recording Standard Hub.
A legal opinion cited in the Oct. 12 reporting argues the failure to inform applicants about AI use is likely unlawful, flagging data protection and human rights concerns — an attributed claim, not a court finding. Asylum seekers aren't told when AI is used on their claims, per the reporting. What's still unestablished: the volume of affected cases, the rate of harmful errors in real decisions, and whether any individual decision has been overturned as a result.
One thing worth keeping straight: UK border facial recognition for estimating asylum-seeker ages is a separate system. In June 2026 the Home Office announced a rollout, and on June 19 more than 60 international human rights groups demanded it be cancelled, warning of bias against women and people of colour. The tool is meant to classify arrivals as under or over 18; a separate ORG-related line of that debate concerns performance in the 16–18 window, where misclassification carries the hardest consequences. That pipeline takes an image and returns an age estimate. ACS takes a transcript and returns a summary. Different input, different failure mode — which is exactly why the ACS numbers should be read on their own.
The decisive unknown stays where it was: whether Home Office reviewers compared ACS summaries against source transcripts before decisions, and at what rate.
Comments ()