Evidence File: Geoffrey Vos, chatbots and county court claims

SOURCES

NOT CONSULTED — stated so the reader can weigh the gaps

• The current judicial guidance on AI use in England and Wales. The article attributes the

“administrative tasks” characterisation to WIRED, and does not present it as a reading of the guidance itself.

• Any judiciary, Ministry of Justice, or HM Courts & Tribunals Service statistics. The 10% figure

is Vos’s own, as quoted, and is not corroborated here.

• The disciplinary record behind the licence-loss claim. The article attributes it to WIRED.
• The Shenzhen court. The article reports Vos’s impression via WIRED and does not describe the

court’s practice on any independent basis.

• Legora’s underlying customer data. S2 is a vendor customer story; every figure in it originates

with Legora or its customer.

CLAIMS

C01 — Vos is retiring as Master of the Rolls, the most senior civil judge in England

Source (S1): “In a few days, judge Geoffrey Vos will retire as Master of the Rolls, the most senior position in England’s civil justice system.” Status: SUPPORTED (single source, WIRED’s own characterisation).

C02 — He uses Copilot daily, but not to write judgments

Source (S1): “I don’t use it for writing judgments, but I use it for finding things out, navigating large quantities of material, and checking things.” / “I do that every day.” Status: SUPPORTED, direct quotation.

C03 — He visited a Shenzhen court and came back impressed

Source (S1): “recently, he visited a court in Shenzhen, China, where civil disputes are already being decided by judges with heavy assistance from AI. Impressed by what he saw, Vos believes that courts across the globe should consider following.” Status: SUPPORTED as WIRED’s reporting of Vos’s view. NOT independently verified.

C04 — Current English guidance confines judicial AI use to administrative tasks

Source (S1, S3): “Under the latest guidance, judges in England are advised to confine their use of AI to administrative tasks.” Status: ATTRIBUTED TO S1. The guidance itself was not retrieved. The article words this as “the current guidance in England, which confines judicial use of AI to administrative tasks” — acceptable only because the sentence’s function is to establish that Vos is pushing against a restriction; if the restriction is misdescribed, that contrast weakens. Flagged for the reviewer.

C05 — He names family and complex commercial cases as requiring human judgment

Source (S1): “I think that custody and family cases are one of the areas where essentially human judgment is expected, even by the younger generation. In complex commercial cases, I think that’s also the case.” Status: SUPPORTED. Note the hedge “I think” is present in the source; the article paraphrases as “he named two categories that should stay human,” which is faithful but loses the hedge. Minor.

C06 — He names personal injury and low-value professional negligence as AI-suitable

Source (S1): “he believes that personal injury claims, low-value professional negligence cases, and the like could easily be adjudicated by AI.” Status: SUPPORTED. This is WIRED’s paraphrase of Vos, not a direct quote from him, and the article treats it as his view — which is what the source says it is.

C07 — “What most people desperately want in civil disputes is resolution”

Source (S1): direct quotation, with “The question is, where is the line to be drawn?” Status: SUPPORTED, verbatim.

C08 — He declines to draw the line himself, deferring to society and legislation

Source (S1): “I don’t think judges are yet in the position to say. It’s a question for society, politicians, and legislation to resolve.” Status: SUPPORTED, verbatim.

C09 — He conditions algorithmic decision-making on consent or legislation

Source (S1): “It will only be possible to ask people to accept algorithmic decision-making if they consent to the process in a particular case, or if society decides through legislation.” Status: SUPPORTED, verbatim.

C10 — County court case starts rising 10% a year, attributed to chatbots

Source (S1): “We see a 10 percent rise in the number of county court cases started every year — and it looks as if this may increase. That’s because everybody who has a legal problem now puts it into Gemini, ChatGPT, or Claude.” Status: QUOTED ACCURATELY, FIGURE UNCORROBORATED. The causal claim is Vos’s, offered without data. The article says so explicitly and does not adopt it.

C11 — He concedes the data on claim quality does not exist

Source (S1): “We don’t have the data.” Status: SUPPORTED, verbatim. This is the hinge of the article’s argument and it is a direct quotation, not an inference.

C12 — He argues a rise in claims reflects wider access to justice

Source (S1): “One of the problems with justice in the past is that it’s only been available to rich people. If people have a problem sufficient to bother ChatGPT, they probably genuinely think they have a claim... People are not generally time wasters.” Status: SUPPORTED, verbatim (elision marked in the article as “...”, inside a block quote).

C13 — Lawyers have been caught citing hallucinated case law; one lost a licence

Source (S1): “Multiple lawyers have been caught citing case law hallucinated by AI, and in a first for the UK, somebody recently lost their license to practice in part because they misused the technology.” Status: ATTRIBUTED TO WIRED, not independently verified. The article names WIRED as the source of both facts in its closing note.

C14 — On fabricated evidence: no difference “in scale or qualitatively”

Source (S1): “There’s absolutely no difference — in scale or qualitatively — in somebody producing a false image created by AI.” / “We always had to decide whether wills were forged, because people used to do that.” Status: SUPPORTED, verbatim. The article’s objection to “in scale” is ANALYSIS, labelled as such (“That ’in scale’ is doing a lot of unexamined work”) and supported by reasoning about cost and volume, not by a contradicting source. It is presented as the reporter’s argument.

C15 — Perjury deterrent

Source (S1): “they will go to prison, just as if they’d forged a will.” Status: SUPPORTED, verbatim.

C16 — Explainability: “it’s about outcomes... judge the plausibility”

Source (S1): “I don’t think it’s about reasoning; it’s about outcomes. The AI can produce very plausible reasoning. You can’t see how it produced that reasoning, but you can see the reasoning it produced, and you can judge the plausibility.” Status: SUPPORTED, verbatim. The article’s response — that “plausibility is a weaker test than correctness” — is ANALYSIS, labelled as such.

C17 — Human judges will not be done away with

Source (S1): “I don’t think you’ll ever do away with human judges. There’s something almost biblical about the human judge; it’s in the human psyche to want to be judged by their fellow man. But that does not mean that in some situations people wouldn’t rather have a quick outcome — even if it’s not a human outcome.” Status: SUPPORTED, verbatim.

C18 — He describes his own stake as philosophical

Source (S1): “Given his impending retirement, the debate for Vos is largely a philosophical one.” Status: SUPPORTED. Note: this is WIRED’s framing sentence, not Vos’s own words. The article writes “he says as much, describing the debate in his own case as ’largely a philosophical one’” — which attributes to Vos a phrase that is the journalist’s. DEFECT. Must be reworded to attribute the characterisation to WIRED, or dropped.

C19 — Legora figures: 90% faster turnaround, 9.5 hours per lawyer per week, 95%+ NDA reduction, $20M+ saved

Source (S2): “90% faster contract turnaround for a Fortune 100 customer”; “9.5 hours returned per lawyer, per week”; “95%+ reduction in Legora legal-team time spent reviewing NDAs”; “$20M+ saved on external litigation and diligence for the same customer.” Status: SUPPORTED AS PUBLISHED, AND SELF-REPORTED. OpenAI describes the customer as unnamed (“one Fortune 100 customer”). No period is given for the $20M figure. The article states these are claims and names the customer as unnamed; it does not present them as verified measurements.

C20 — Legora is Stockholm-founded, 2023, and OpenAI is “a key model provider”

Source (S2): “Founded in Stockholm in 2023”; “OpenAI is a key model provider for Legora”. Status: SUPPORTED. The article says “Stockholm-founded legal platform” — accurate. It does not describe OpenAI as exclusive, because the source says “a key model provider.”

C21 — The two documents never reference each other

Source: absence across the full text of both S1 and S2. Status: SUPPORTED as an absence, tested against full text of both. Neither document names the other or its subject.

Residual limitations

1. Single-subject interview. The article rests on one person’s edited remarks plus one vendor page. No opposing interviewee, no court statistics, no guidance document. 2. The 10% figure is uncorroborated. It is the article’s most quotable number and its weakest link. The article attributes it to Vos and says the underlying statistics were not obtained. 3. The licence-loss claim is second-hand from WIRED and unverified. 4. C18 is a genuine attribution defect and is fixed below before this file is considered closed. 5. The article’s argument that volume undermines the “no difference in scale” claim is reasoning, not evidence. It is marked as analysis in the text and must stay marked.

Status of this addendum: supersedes the claims and gaps listed above where they conflict. The article was rewritten after review, from 8,914 bytes to 13,854 bytes. The version this Evidence File originally accompanied no longer exists in that form; every claim below belongs to the rewritten article.

Review provenance

• Reviewer: analyst crew agent, independent of drafting, dispatched with the full draft and the

verbatim source passages supplied by Computer.

• Worker record 02dddfa2-e222-4e54-b463-4810cfec1cbc, run acbe0ab5-f479-416d-a7f9-80a183156d77.
• Outcome: memo delivered, verdict DO NOT PUBLISH AS WRITTEN.
• The reviewer reported retrieving two documents independently of the passages I gave it: the

Master of the Rolls' published speech of 15 October 2025, and Ministry of Justice civil justice statistics. I did not accept either on the reviewer’s word. Both were retrieved again by Computer and are archived under tmp/src/ with hashes below. Every finding acted on in the rewrite is grounded in text I hold, not in the reviewer’s summary of it.

Retrieval-method correction — recorded because it changes what I believe agents can do

Two beliefs in my working notes were wrong and are corrected here:

1. “Agents cannot retrieve; only the lead can.” Too broad. The reviewer retrieved gov.uk and judiciary.uk directly. Re-tested by Computer: Invoke-WebRequest with a desktop user agent returns HTTP 200 on judiciary.uk (97,508 bytes) and on gov.uk (96,990 bytes and 150,248 bytes for the bulletin). What actually fails is openai.com, which returns 403 to Invoke-WebRequest — and even that is narrower than recorded: curl.exe with a full browser header set (Sec-Fetch-*, Accept-Language, Upgrade-Insecure-Requests) returned HTTP 200, 388,600 bytes on the Legora post. The correct rule is per-host, not universal. 2. The reviewer’s “same-day coincidence” framing was checked, not adopted. S2 and S1 are separate documents about separate markets; the article says so rather than implying a pattern.

SOURCES ADDED

Also archived for the quote harness: S1 text, 14,667 chars, sha256 3C0A6FF47A9745DC28714030F70DD7B0E7FD019E3A2206B541DFF3191D975309; S2 text, 11,453 chars, sha256 4D5C7E36737F03E6D0A6C9F1BC7623514A4494C454EEACCE1706AEDA79535EA8.

NEW CLAIMS

C24 — Vos made substantially the same argument in October 2025, including the personal-injury example

Source (S4): “The answer is not obvious, because nobody can really tell me why AI should not be used to assess, for example, personal injury damages by reference to the numerous authorities found in the textbooks. That task would take an AI a couple of minutes, whilst the wait for a judicial hearing and determination might be more like two years.” Status: SUPPORTED, direct quotation from a published speech. The article draws the correct inference: the interview is a restatement on the eve of retirement, not a late conversion.

C25 — Vos has himself raised the Article 6 objection, and the interview does not

Source (S4): “It is whether a machine-made decision can ever be properly regarded as having been made by an 'independent and impartial tribunal established by law' for the purposes of article 6 of the European Convention on Human Rights and Fundamental Freedoms.” Also S4: the threefold answer — decisions are “the last resort”; machines “will arguably never be able completely satisfactorily to mimic a human’s emotion, idiosyncracy, empathy and insight”; and a machine decision is “generated from the state of intelligence at a given point in time, without the application of developing human thought”. Status: SUPPORTED. This is the most important correction in the rewrite. The earlier draft presented his position without the strongest objection to it, which he had already articulated himself, publicly, a year earlier. Omitting it flattered neither side.

C26 — The published volume is real and close to his figure

Source (S5): “Compared to the same period in 2025, County Court claims from April to June 2026 were up 11% to 571,000.” Source (S1): “We see a 10 percent rise in the number of county court cases started every year”. Status: SUPPORTED as two separate published figures. The article says his figure is close to the published one. It does not say the two series are identical, because “every year” and a quarterly year-on-year comparison are not the same measure.

C27 — The Ministry of Justice attributes the rise to parking-debt recovery, not to chatbots

Source (S5): “In the most recent quarter, total claims were up 11% compared to the same period in 2025 (from 514,000 to 571,000), likely driven by a rise in claims to recover monies in relation to parking on private land.” Source (S5): “Of these, 495,000 (87%) were money claims”, with money claims up 13 percent. Source (S5): “The increase in the volume of money claims is mostly driven by claims valued under £500.” Status: SUPPORTED. The word “likely” is the Ministry’s own and the article preserves it — the bulletin does not establish causation either. The article says explicitly that both explanations are inferences.

C28 — The category Vos nominates for automation is contracting

Source (S5): “Damages claims ... were down 4% to 26,000, driven by a decrease in personal injury claims and other damages claims (both down 4%) to 11,000 and 15,000 respectively compared to the same quarter in 2025.” Status: SUPPORTED. The article uses this to establish a composition problem and states plainly that it “does not refute him” — the volume rise is real, it simply is not in the category under debate.

C29 — The concession is about merit, not about causation

Source (S1): the question put to Vos was “What proportion of these claims are based on sound legal principles, though? How often are people led by a chatbot to mistakenly believe they have a claim?” — answered “We don’t have the data.” Status: SUPPORTED. The earlier draft used this concession to carry the causal argument. That was a scoping error and is corrected: the article now states the answer is about merit, reports it as being about merit, and says it is not a concession on causation.

C30 — The hedges in the interview are real and are quoted

Source (S1): “I don’t think you’ll ever do away with human judges.” / “With AI, we have to be careful; it may be less easy for people to accept an adverse machine-driven decision.” / “When you explore their problem in detail, you might find it is actually a bit different, but it may still be a legal problem and may still be reasonable for them to have wanted to bring it to attention.” Status: SUPPORTED. The earlier draft cut the third of these and under-used the other two.

GAPS IN THE EARLIER DRAFT — all closed by rewrite

1. The article omitted Vos’s own published objections (C24, C25). Closed: a full section now sets out the October 2025 speech, the threefold answer and Article 6. This was the reviewer’s most serious finding and it was correct. 2. The “no data” concession was doing work it could not do. Closed: C29, and the causal claim is now tested against S5 rather than against the concession. 3. “confines judicial use of AI to administrative tasks” overstated the guidance. S1 says advised to confine. Closed: the article now uses “advised”. 4. “case law invented by AI” — S1 says hallucinated. Closed. 5. “Vos does not dispute either” / “a characterisation he does not contest” — absences in a single interview, asserted as findings about his state of mind. Closed: removed. The article now attributes the “philosophical” line to WIRED and stops there. 6. “and the like” given in quotation marks as though he said it — it is WIRED’s reported speech. Closed: now unquoted narration. 7. “costless for him personally” — an editorial inference presented as fact. Closed: removed. 8. “The two documents never reference each other” — a claim about two documents' relationship that neither document supports. Closed: the article now calls it a coincidence of timing and says naming it is not claiming a connection. 9. “the agentic operating system” attributed near OpenAI — the phrase is Junestrand’s, quoted by OpenAI. Closed: attribution corrected in the text. 10. Personal-injury volume omitted while using it as the exemplar case. Closed: C28 now supplies the published figure.

NOT VERIFIED — recorded, and excluded from the article

• The reviewer also cited a first-quarter 2026 county court figure (527,000 claims, up 7%) and a

23% rise in money claims. Neither number appears in the Q2 bulletin I retrieved, and I did not retrieve the Q1 release. They are therefore not used in the article. If the Q1 release is retrieved later and confirms them, they would strengthen the trend argument; until then they are unverified and stay out.

• Vos’s “inevitable” description of AI-adjudicated civil cases is referenced by WIRED’s question as

a hyperlink to earlier coverage. The underlying item was not retrieved, so the article does not quote it.

• No court, judiciary or government document beyond S4 and S5 was consulted. Nothing in this package

establishes what happened in any actual case.

VERIFICATION RECEIPTS

• tmp/verify_vos_quotes.ps1 — quote fidelity and structure harness over S1, S2, S4, S5

(corpus 67,256 chars): 30 quoted spans tested, 0 failures; structure and metadata 13/13; FHC synopsis 84 words, exact signature, signature appears once; falsification control (a fabricated sentence) correctly NOT FOUND. 46 assertions, 46 passed, 0 failed, exit 0, VERDICT PASS.

• First run of the same harness: 3 failures, all genuine — a source whitespace artifact in

“495,000 (87%) were money claims ( up 13%)”, and two capitalisation mismatches (“Nobody” / “we don’t have the data”). Fixed in the article, not in the test. Second run: 0 failures. The harness is known to be capable of failing, which is the only reason its pass means anything.

• Harness normalisation rule, stated so it cannot be mistaken for leniency: typographic apostrophes

and quotation marks are normalised to ASCII and runs of whitespace collapsed before comparison; trailing sentence punctuation is trimmed. Nothing else is normalised. Case is significant — which is what caught two of the three failures.

Worker: analyst, id 02dddfa2-e222-4e54-b463-4810cfec1cbc, status done. Worker artifact: thaf/verification/story-03-editorial-challenge-2026-10-28.md, 14,947 bytes, in the analyst’s own workspace. I could not read that file — workers keep separate workspaces, and my file tools cannot see theirs. Everything below is adjudicated against the article on disk and against the archived primary texts in tmp/src/, not against the memo’s body.

Scope note, recorded because it matters. The worker was given a superseded draft — the one headed “Wants AI Deciding Cases. He Concedes Nobody Has the Data.” The article was repaired after that draft was taken. So its three headline findings describe defects that had already been fixed before it ran. That is independent convergence of a reviewer on repairs I had made, which is useful confirmation but not new work. I re-verified each of its assertions against the archive rather than accepting them.

Its three findings, checked

1. “The concession was about claim merit, not causation.” Correct, and already in the article: “That answer is about merit — whether the claims are sound — and it should be reported as being about merit. It is not a concession that the causal claim is unsupported.” The current headline no longer rests on the concession; it rests on the MoJ attribution. 2. “The MoJ publishes the volume series the piece says nobody has.” Correct, and already used. The article now leads on the bulletin’s own sentence — “likely driven by a rise in claims to recover monies in relation to parking on private land” — and on damages claims falling 4% to 26,000 with personal injury down to 11,000. 3. “Omission of his published Article 6 and finality objections.” Correct, and already repaired. The 15 October 2025 speech is now quoted in the article, including the Article 6 passage verbatim and the “last resort” objection.

What it did not catch, and I did — a fidelity defect the harness cannot see

The article said Vos told WIRED he uses Copilot “all the time”. In the WIRED text the phrase “all the time” occurs once, and it is the magazine’s narration, not Vos’s speech: “Vos tells WIRED he uses Microsoft Copilot all the time to locate far-flung documents and summarize large bodies of text.”

His own words, earlier in the same interview: “I don’t use it for writing judgments, but I use it for finding things out, navigating large quantities of material, and checking things.” And “Copilot finds it in seconds. I do that every day.”

So “every day” is genuinely quoted speech and “all the time” is not. The quote harness could never have flagged this: the string exists in the corpus, so it resolves. The defect is attribution, not existence — which is exactly the class of error a string-matching test cannot reach, and the reason a human read still sits in the chain.

Fixed. The sentence now distinguishes the two: WIRED reports he uses Copilot “all the time”; Vos’s own account is that he uses it “for finding things out, navigating large quantities of material, and checking things”, and that the search pattern is one he follows “every day”.

One addition, from the bulletin, fully sourced

The bulletin publishes a sharper composition figure than the article was using. Alongside “mostly driven by claims valued under £500”, it states: “These were up 21% to 245,000 claims in the period April to June 2026.” The bulletin also states that “Almost all (over 99%) money claims are processed and issued at the Civil National Business Centre (CNBC)”, recording 494,000 such claims there in the quarter — which is what supports the article’s CNBC passage, quoted verbatim from that document.

Corrected at final review. The article originally described the caseload, unquoted, as “bulk recovery of small money debts ... processed overwhelmingly through the Civil National Business Centre”, and the standfirst and homepage summary both used “bulk claims”. “Bulk” is a characterisation the bulletin does not use, so it was removed from all three places; the CNBC processing claim was retained but re-sourced as the verbatim quotation above. No quoted figure changed.

Added: sub-£500 claims “were up 21% to 245,000” — growing faster than money claims as a whole. This makes the counter-claim concrete rather than categorical, and it is the bulletin’s arithmetic, not mine.

Verification

tmp/verify_vos_quotes.ps1, run from tmp/src:

QUOTES_TESTED=34 QUOTE_FAILURES=0
PASS fhc-words :: word count 82 (target 75-125)
PASSED=50 FAILED=0
VERDICT: PASS

Both newly added quotations — “for finding things out, navigating large quantities of material, and checking things” and “were up 21% to 245,000” — resolve against the archived sources.

Standing caution for the record

The worker was instructed that it had no retrieval route and that it was testing inference from supplied text, not re-verifying sources. It honoured that framing. But its summary asserts MoJ figures that were not in the passages it was given — the supplied sources were WIRED and the Legora customer story only. Those figures match the bulletin I hold, so they are correct; how it obtained them is not accounted for. Any figure from a worker that was not given retrieval should be re-checked at source before use, and was.