Approval without engagement is counterfeit.

I chose the word counterfeit deliberately, and I have been asked more than once whether it is too strong. I think it is exactly the right strength, because counterfeiting is not a failure of skill. It is passing off a thing as something it is not, and looking convincing while you do it.

Here is the scenario, and every practitioner reading this recognises it because it is the middle of a bad week rather than a hypothetical. A draft comes back from the machine. It reads well. It has the right shape, the right headings, the confident tone of somebody who knows what they are talking about. There are eleven other matters live and a hearing tomorrow. You skim it, you cannot see anything obviously wrong, and you approve it. Nobody lied. Nobody was reckless in any way they would recognise in themselves. And the work has gone out under a professional's name without a professional ever having formed a view about it.

That is not supervision. Supervision is not a state of having looked; it is the act of being capable of disagreeing and having exercised the capability. What happened in that scenario is that the signature was applied without the thing the signature represents. The client received a document bearing all the marks of professed advice, and the marks were not backed by the reserve.

The courts have already met the consequences, and they have not been subtle about it. The reason the Bar Standards Board issued dedicated guidance on the use of artificial intelligence and other technologies, in force from 18 May 2026, was in part the profession's encounters with fabricated citations, in Ayinde v London Borough of Haringey among others. Fake authorities in submissions. Cases that do not exist, cited to a judge, by counsel. Nobody involved set out to mislead a court; every one of them approved something they had not checked. The distance between those two states is one afternoon of pressure, and the second one is a career event.

The relevant rule is not new and was not written with models in mind. Paragraph 1.4 of the SRA Code says you do not mislead or attempt to mislead your clients, the court or others, either by your own acts or omissions or allowing or being complicit in the acts or omissions of others. Read the omissions limb slowly. Passing unexamined machine output to a client or a court is complicity by omission, and it is complicity whether or not the output turns out to be wrong; you were simply lucky in the cases where it held. On the Bar side rC3.1 prohibits knowingly or recklessly misleading the court, and guidance gC65 closes the loop by confirming that where tasks are delegated or outsourced you remain personally responsible for the work. The SRA said the same in terms in its Risk Outlook report on artificial intelligence back in November 2023: you remain responsible and accountable for the outputs from AI you are using, and however well controlled a system is, firms will still need to check its outputs for accuracy.

Which brings me to the sentence in this declaration that I care about most. AI errors in law are supervision failures, not technology failures.

I have held that view for a couple of years now and it is not a rhetorical flourish, it is a diagnosis with practical consequences. If you treat a hallucinated citation as a technology failure, your response is to procure a better model, wait for the vendor to fix it, and carry on with the same workflow. If you treat it as a supervision failure, your response is to look at your own process and ask why it allowed unverified material to reach a client. The first response outsources the problem to a company that will accept no liability for it. The second one keeps the problem where the regulator is going to put it anyway, which is on the desk of the professed human whose name went on the outcome.

And it changes what you build. If supervision is where failure actually happens, then supervision is the thing to engineer, rather than the thing you bolt on afterwards in the form of a checkbox and a hopeful policy. That means review designed into the workflow at the points where it is load bearing; verification of anything citable, always, without exception, because that is the failure mode the courts have already seen; a record of what was checked and what was changed; and a culture where saying I have not read that yet is unremarkable rather than an admission of weakness.

There is a harder cultural piece underneath it. The pressure to approve quickly is real, and it is generated by the same economics that make the machine attractive in the first place. Any firm that adopts AI without deciding, explicitly, how many hours of genuine scrutiny it is prepared to fund has already made the decision by default.

My name at the bottom of a document is a representation that I formed a view. On the day I let it mean anything less, it is not a signature. It is a forgery of my own hand.