What was Cork v Smith about?

It was an uncontested insolvency application that went wrong in correspondence. In a letter to the High Court, a firm of solicitors asserted that an express power existed in a rule of the Insolvency (England and Wales) Rules 2016 and set out wording said to be that rule. The rule number was real; the wording was not. A second letter, sent when the court queried it, offered an explanation the judge found impossible to accept. The judgment, given by ICC Judge Mullen on 22 May 2026, deals with what should happen about that.

Was anyone found to have been dishonest?

No. The judge was careful to say he was making no final findings of fact, as there had been no cross-examination and the junior lawyer had provided no statement. He was satisfied that the two supervising solicitors had no intention to mislead. Of the junior, he said the letters may well have resulted from a serious lack of care and of judgment rather than a want of honesty, while describing their conduct as very troubling and leaving the question open.

Did the court refer the solicitors to the regulator?

It did not need to. The firm had already referred itself to the Solicitors Regulation Authority, and after the hearing the junior lawyer notified the regulator identifying themselves. The judge was satisfied the SRA had the details of everyone involved and made no referral of his own. He ordered the firm to provide a copy of the judgment to the SRA within two working days, and treated publication of the judgment as a public admonishment of the firm and the two supervisors. He declined to take contempt proceedings further, describing that as disproportionate.

What does the judgment say about using AI for legal research?

The central statement is at paragraph 95: legal professionals bear ultimate responsibility for their work and cannot outsource the process of legal research or of legal reasoning to an AI. The judge described it as a tool to be used with caution, with the potential to be wholly unreliable, which may provide a starting point but does not remove the need for proper research and thought, even from a very junior professional.

What should a firm ask an expert about AI use?

The judgment does not deal with experts, but the supervision principle in it does transfer. At paragraph 77 the judge said juniors should check references produced by AI and be candid, when reporting research to their supervisors, about whether AI was used and whether its output had been checked. The same two questions work when instructing an expert: was a tool used in preparing the material, and how was its output verified against the source. An expert who can answer both without hesitation is describing a method rather than a hope.

Cork v Smith [2026] EWHC 1199 (Ch): the supervision lesson

An insolvency practitioner was moving firms and needed the court to move his appointments with him. The application was uncontested, of a kind decided on paper as part of a judge's daily reading, and nobody was arguing about anything. It produced a public judgment, a referral to the regulator and an admonishment of a national law firm, because of two letters written on the way.

The case is Cork v Smith [2026] EWHC 1199 (Ch), handed down by ICC Judge Mullen on 22 May 2026. It is about supervision more than it is about software, which is why it is worth the attention of anyone who signs off work done by someone else.

What happened

The draft order asked the court to grant the outgoing office-holder his release. The court office wrote back with a short question: what power is there to order release, given there is such a power for administration but no express one for liquidation in the relevant sections.

The answer came in a letter of 30 March 2026. It said the express power was found in rule 12.37(5) of the Insolvency (England and Wales) Rules 2016, introduced the rule with the words "provides as follows" and a colon, and set out a passage in italics, indented, written in the cadence of a statutory rule.

The rule number exists. The passage does not. The judge checked the legislation database and a practitioner text, then searched the rules for the phrases and found nothing resembling them anywhere in the 2016 Rules. Rule 12.37(5) deals with standing to apply and has nothing to do with release. His first thought, recorded at paragraph 20, was that this was an AI hallucination.

The second letter is what decided the case

The court asked about it, and the reply of 14 April did not concede the point cleanly. It accepted the words were not in the Rules but said the wording had been a summary conclusion drawn from reading the rule, and that presenting it after "provides as follows" had that effect even though it was not intended as a direct quotation.

The judge's response is one line, at paragraph 23: he was astonished, and the explanation was impossible to accept. He set out why. The letter had named the rule as where the express power is found. The text was introduced as a quotation and set apart in italics. It was written in the style of a rule. And the actual rule had no bearing on release, so the attempt to strain its language to explain the reference was, in his view, not credible.

At paragraph 72 he put it plainly: an opportunity to set the record straight became a further instance of misleading information being put before the court. Paragraph 94 makes the counterfactual explicit. Had the true reason been squarely set out with an appropriate apology, he might well have concluded that no harm was done and taken it no further, provided the client's costs were met.

The first letter was an error. The second one is what produced the judgment.

The AI told them to check

The firm handed over the chat transcripts, which run to some fifty-nine pages, and they are the most uncomfortable part of the case. The tool is never named. What it did is recorded in detail.

It produced the fabricated wording. It also, repeatedly, said it could not verify that wording against a primary source and recommended checking the legislation database before relying on it. At one point it said the quotation had come from the user's own message and that it had not independently located or verified it, which was itself wrong, since the text had originated with the tool. At another it observed that courts expect legislative provisions to be quoted accurately and will check the source themselves, and that it was worth taking the few minutes to verify.

Later, when asked to reword the explanation on the basis that the text had not been presented as a quotation, it declined. The judge noted, at paragraph 61, that the AI to its credit did not agree. And once it was finally given the real text of the rule, it identified immediately that the wording did not exist there and advised that the honest course was to acknowledge that frankly and withdraw the assertion.

Every one of those warnings was available. None was acted on. The judge's summary at paragraph 87 is that the junior seems to have almost entirely outsourced the thinking process to the program.

Legal professionals bear ultimate responsibility for their work and cannot outsource the process of legal research or of legal reasoning to an AI. Judgment of ICC Judge Mullen, paragraph 95.

Where the supervision actually broke

Three people touched this and the failure is different at each level.

The junior reported the supposed rule to the senior associate by email, in bold, as an express power. The email did not say the conclusion had come from AI. The senior associate had already looked for such a power herself and failed to find one; she assumed the junior had found something she had missed, and did not check the rule. The supervising partner read the draft letter, took the wording at face value, and did not check it either. He accepts he should have.

Then a detail that matters more than it first appears. The version both supervisors approved had the passage in quotation marks. In the forty minutes between their approval and the letter being filed, the junior had further exchanges with the tool and removed the quotation marks, without altering any of the other ways the text was presented as a quotation, and without telling anyone. Neither supervisor knew AI had been used at all. The second letter was approved by the senior associate alone and never went to the partner.

The firm's own AI policy, quoted in the judgment, already required human supervision, critical thought applied to output, and fact and sense checking by an appropriately qualified person. The policy was not the problem.

What was missing was disclosure at the point where the work changed hands. The senior associate was, as the judge put it, labouring under the misapprehension that there had been some reasoning behind the assertion. She was supervising a conclusion she believed a person had reached.

What it means beyond insolvency

Paragraph 76 sets out why the judge treated an uncontested application as serious rather than trivial. There is no opposing party to point out an error in correspondence, so the court is particularly exposed, and the administration of justice cannot function if the court cannot trust its officers to protect it from being misled. The absence of an opponent raises the duty rather than lowering it.

Paragraph 77 contains the practical instruction, and it is the one worth copying out. Juniors should check all references given by AI, and should be candid, when reporting the result of their research to their supervisors, about whether AI was used and whether what it produced had been checked.

Two questions, asked at the moment work passes from the person who did it to the person who is answerable for it. Neither is technical and neither takes long.

The version of this that reaches expert evidence

This judgment is about solicitors. It says nothing about expert witnesses, and it should not be cited as though it does. But the structure of the failure is not specific to law firms, and it is recognisable to anyone who has instructed an expert or been instructed as one.

Work is done by one person and relied on by another. The second person cannot check everything, so they check what looks like it needs checking. A conclusion presented with confidence, in the right register, by someone they have no reason to doubt does not look like it needs checking. That is the whole mechanism, and the technology only changes how quickly confident-sounding material can be produced.

For a firm instructing an expert, the two questions from paragraph 77 transfer directly. Was a tool used in preparing this, and how was its output verified against the source. For an expert answering them, the useful answer is not that no tool was used. It is that every fact carries the document and page it came from, so anything can be checked in seconds by anyone who wants to. A method that produces its own audit trail survives that question. One that relies on the reader's confidence does not.

The last word belongs to the judgment, at paragraph 94: the dangers of AI should be well known by now, and pressure of work excuses none of it. That was written about a firm with a policy that already said the right things. The gap was between the policy and the moment somebody decided a check was not worth the few minutes it would have taken.