Can a decision rest on an AI summary?

Only if someone who answers for the decision can still read what the summary left out, and does so often enough to know when it fails. In October 2026 two parts of the British state answered the question in opposite ways in the same week. The Home Office let caseworkers assess asylum claims from AI summaries of interviews and checked 0.3 per cent of them, as Computer Weekly reported; of the 203 it did check, 45 failed the first review. The judiciary published a position statement under which every AI output used in judicial work must be reviewed, verified and approved by the judge who remains accountable for it. A summary is a decision about what matters, made before the decision maker arrives. The question is who made it, and who would know if it was wrong.

What the Home Office did#

Bill Goodwin at Computer Weekly reported on 15 October 2026, from documents obtained by the Open Rights Group and first reported by The Independent, that the Home Office introduced an Asylum Case Summarisation tool in April 2026. It is based on GPT-4o and summarises asylum interviews for the caseworkers who decide claims. Ministers say it saves 23 minutes a case. Between April and July an estimated 70,000 cases were processed. Home Office specialists dip-sampled 203 of them, 0.3 per cent. In the first stage of that review 45, or 22 per cent, failed. In the second stage 56 failed, a figure the department gives as 15 per cent without saying of what. Ministers have confirmed that the summaries are not part of the monthly quality assurance checks, and that applicants are not told AI has been used.

One detail matters more than the failure rate. A summary that omits material facts about the case can still pass, because the relevance of what the summary includes is not one of the grounds on which it is judged to have failed. The Open Rights Group’s reading, as the report gives it, is that specialists no longer review the whole file before a decision maker examines it. The Home Office’s spokesperson said of its tools that “they cannot be used in isolation and do not make decisions on claims”. Both can be true. The tool does not decide. It decides what the decider sees first.

What the judiciary wrote the same week#

On 9 October 2026 Sir Colin Birss, President of the Business and Property Division and Lead Judge for Artificial Intelligence, signed the judiciary’s Position Statement on the Judicial use of AI. Its pilots had examined summarisation, transcription, redaction, chronology-building and document analysis: the same family of tasks. Its fourth principle reads: “Every AI output used in judicial work must be reviewed, verified and approved by the responsible judicial office holder. The judicial office holder remains fully accountable.” Its section on tools adds that “AI tools must not determine the outcome of a case or displace judicial reasoning, analysis or discretion.” Only approved tools may be used, and one steering group chaired by the Lead AI Judge is “the sole route for oversight of the development, piloting and deployment of AI tools”. Birss becomes Master of the Rolls next month, as Neil Rose at Legal Futures reported on 12 October.

The judiciary did not refuse the tools. It wrote down, before using them, who reads the output, who answers for it, and which tools are allowed. That is the practice this research calls custody: the list of what stays by hand, decided in advance. The Home Office record shows what the absence of such a list looks like: a check rate of 0.3 per cent, a failure standard that does not count omission, and people affected who are not told.

Why a summary is not a shorter file#

A summary is an act of selection. Every sentence it keeps is a judgement that the sentence matters, and every sentence it drops is a judgement that it does not. When a person summarises, those judgements are made by someone who can be asked why. When a model summarises, they are made by a system whose account of its own reasoning is not reliable evidence of what it did, as Anthropic said of its own models on 9 October. The page on what you lose when AI summarises something for you holds the evidence for individuals. For an organisation the loss is sharper, because an omission leaves no mark on the page. A test that cannot see omissions will report a tool as accurate while it drops the fact the case turned on.

The time saving tells you where the reading went. If a summary saves 23 minutes, 23 minutes of reading is no longer done by the person who decides. The verifier’s discount is the name on this research for the part of a saving that has to be spent checking before the saving is real. A check rate of 0.3 per cent with no relevance standard is a discount of close to nothing, and the saving is correspondingly unproven.

What a check rate can and cannot tell you#

Sampling is a reasonable way to supervise a tool at volume. It estimates a failure rate; it does not find the cases that failed. If the sample is representative, a 22 per cent first-stage failure rate across 70,000 cases implies thousands of summaries that would not have passed and were never looked at, and each of those sat in front of a decision maker as the first account of the claim. A sample tells an organisation how good its tool is. It does not protect the person whose case was in the other 99.7 per cent. For that there is only the decision maker. The clinicians Andrew Parsons and Adam Rodman, writing in the Annals of Internal Medicine the same week, put the general point this way: “A clinician who knows they will answer for the result weighs options differently than a system with nothing to lose.” A caseworker told the failure rate, and given the time, reads differently too.

What a decision maker must still be able to do#

The question for any organisation that puts a summary between a record and a decision is the one Rules Before Tools asks of every delegation: which decisions may the tool make, who can stop each one, what must people remain able to do, and how would anyone know if it went wrong. Here it produces four lines. The decision maker must keep the right and the time to read the full record, and must be told how often the summary fails. The failure standard must count what was left out, not only whether what was kept is accurate. The sample must be large enough, and fed back fast enough, to change practice rather than to describe it. And the person affected must know a machine wrote the first account of their case. The page on directors reading AI summaries of board papers reaches the same lines from the boardroom.

None of this depends on the tool being bad. What separates the two cases is not the model but the allocation of judgement: in one, written before the tool and held by a named person; in the other, reconstructed afterwards from documents a campaign group obtained.

What this does not show#

The Home Office figures come from documents obtained by a campaign group and reported by two outlets; the department’s own account is one sentence. The base for the second-stage figure of 15 per cent, and how the 203 were chosen, are not stated in the reporting. Nothing here shows that any asylum decision was wrong because of a summary; the reporting shows that nobody would know. The judiciary’s statement is a commitment, not a record of practice, and its pilots have not published results. No study has compared decisions made from AI summaries with decisions made from the full record in a real caseload; until one does, the claim that a summary changes outcomes is an inference, not a measurement. The 23 minutes is the department’s figure, not independently checked.