Devilling for the Machine: rC86, AI Outsourcing and the Verification Problem

A scan of the senior bar's Legal 500 profiles produces a remarkably uniform vocabulary: laser-sharp minds, brains the size of a planet, "intellectual rigour" deployed without irony. Client care gets a polite paragraph; the cerebral self-portrait gets the rest. None of this is new. What is new is the quiet doubt now sitting behind it: that the analytical feats on which that self-image rests can be matched, and in some narrow registers exceeded, by software running on a laptop.

That doubt arrived in court this spring.

At a directions hearing in the Divisional Court, counsel was asked to produce hard copies of two authorities cited in his skeleton. He paused, clicked, and conceded the cases did not exist. They had been generated by a public language model and inserted without verification. The collapse was not unique; it has happened often enough now to count as a sub-genre of judgment. But it captures the contradiction that defines the profession's relationship with these tools: the same software that drafts your skeleton at three in the morning will, with equal fluency, invent the cases it cites.

The legal-tech sector sells the inverse proposition. Because a model can produce a polished document, the pitch runs, it has automated the reasoning behind it. The polish does most of the persuasive work. A well-laid-out submission with proper headings and confident topic sentences passes for thought because it shares the surface signals; the harder question, whether the structure tracks an actual chain of legal reasoning, is the question the model cannot answer about its own output, and which the busy practitioner is increasingly tempted not to ask.

Daron Acemoglu's diagnosis of automation generally is that development is biased toward cost-cutting rather than capability: "so-so technology", in his phrase, which displaces the human worker without producing offsetting gains. That is one register. The competing register is harder to dismiss. OpenAI's research models have, in recent months, produced novel results in combinatorics that took specialist mathematicians by surprise. Both things are true. The technology is genuinely capable and structurally over-sold, sometimes in the same product.

The Bar Standards Board's AI Guidance, issued on 18 May 2026, takes the predictable but useful step of treating generative AI as outsourcing within the meaning of rC86. The analogy is to devilling. A senior barrister who instructs an unseen junior to draft remains personally answerable for the work signed in their name. Core Duty 7 does not soften because the unseen junior is silicon. Rule rC20 does not admit a "the machine did it" plea. The reframing is conservative and exact: it does not impose new obligations so much as make plain that the existing ones already attach.

There is nothing reassuring about this for the practitioner.

In mathematics, the response to AI-generated reasoning is straightforward. The output is piped into a proof assistant (Lean, or Coq) which type-checks each step against the axioms. The mathematics either compiles or it does not. Law has no equivalent. A legal argument's correctness is not a property of its form; it is a property of its fit with the authority it invokes, the facts it engages, and the doctrinal field it sits within. The judge's mind is the only compiler our submissions will ever meet, and a judge under pressure may not catch a confidently mis-stated proposition any more reliably than counsel did.

The judicial response has been firm and now amounts to a small, coherent line. In R (Ayinde) v London Borough of Haringey [2025] EWHC 1383 (Admin) the High Court warned of contempt and regulatory referral for the submission of unverified AI-generated citations. Ndaryiyumvire v Birmingham City University [2025] 10 WLUK 719 produced a wasted-costs order against a firm that filed a draft built around fictitious authority. R (Munir) v Secretary of State for the Home Department [2026] UKUT 81 (IAC) extended the warning to confidentiality, observing that uploading client material to a public model is liable to waive privilege. None of these decisions is doctrinally novel. They are reminders, delivered in increasingly sharp form, that the existing rules cover the existing problem.

I have spent the past year using these tools intensively in my own practice, and the experience has not left me with the easy humanist conclusion. The models are not stochastic parrots. Trained on decades of skilled output, they routinely produce synthesis that a careful single practitioner would not generate alone within the time they have. The trouble is the failure mode. Where a junior tenant who does not know an answer says so, a language model will, with equal fluency, produce the right answer or an invented one, and a reader cannot reliably distinguish them at the surface. The polish is the problem.

The BSB Guidance solves none of this. What it does is allocate the residual risk, accurately, to the person who signs the document. The barrister who circulates an AI-drafted note as their own work has not delegated anything in any sense the regulator recognises; in the BSB's terms they have written it, regardless of how it was drafted. The practical question is no longer whether to use these tools, since most practitioners already do, but how to build the verification layer that the proof assistants do not give us. That work is unglamorous. It is reading every cited case. It is checking that the case stands for the proposition asserted. It is the part of the job that AI does not, in the relevant sense, do.

The machine drafts; the barrister signs; the duty has not moved an inch.

Table of Authorities