Lawyers already know how to instruct precisely

Most AI training for lawyers opens by treating the audience as beginners at something. That framing is wrong, and it costs the room about half an hour of goodwill in the first ten minutes.

Consider what a commercial lawyer does when drafting a clause.

They specify a desired behaviour in natural language. They do it for a reader who is adversarial and will exploit any ambiguity. They anticipate edge cases and handle them explicitly. They define terms to constrain interpretation. They order provisions so earlier ones govern later ones. They add carve-outs for cases the general rule would catch wrongly. Then somebody else reads it looking for holes, and the draft is revised.

That is specification under adversarial reading. It is the same discipline that produces a reliable instruction to a model, and lawyers have spent their careers doing it under conditions where getting it wrong has consequences.

The starting point is further along than most training assumes.

What transfers

Precision about scope. A lawyer who has argued about whether “reasonable endeavours” differs from “best endeavours” already understands that the apparent meaning of a phrase is not the operative one. That instinct is exactly what is needed when an instruction says “summarise the key points” and produces five different notions of “key” across five documents.

Defining terms. Legal drafting handles ambiguity by fixing meaning up front. The same move works on a model: state what a “material” obligation is for this task rather than assuming a shared definition.

Anticipating the edge case. Lawyers do this reflexively — what if the counterparty is in administration, what if notice is served on a Sunday. The equivalent question is what the system should do when the document is a scan, when the clause is absent, when the answer is genuinely unclear. Most bad outputs are edge cases the instruction never addressed.

Structure as control. Numbered provisions, defined terms, ordered operation. The same structure makes an instruction reliable, and for the same reason: it removes discretion at the points where discretion is expensive.

Reading adversarially. The habit of looking at a document and asking how it could be read against you is precisely the habit needed to evaluate an output. It is the single most valuable thing this audience already has.

What does not transfer

Three things, and being honest about them is what earns the rest.

A contract is read by someone who can be held to it. A model cannot. A counterparty who reads a clause perversely can be argued with, and the argument has a forum. A model that reads an instruction unhelpfully just produces the unhelpful output, at scale, silently. There is no forum. This means instructions need to be tested empirically in a way contracts are not — you find out what an instruction actually does by running it across twenty documents and reading all twenty, not by reasoning about what it should do.

Legal drafting assumes shared background; models have no matter context. When a lawyer writes to another lawyer, an enormous amount is carried by context — the deal, the sector, the prior correspondence. A model has the words in front of it and nothing else. Instructions that would be insultingly explicit between professionals are frequently necessary here.

Fluency is not a signal of correctness, and legal training points the wrong way here. A well-drafted document usually indicates a competent drafter, so lawyers reasonably treat polish as weak evidence of care. That heuristic inverts. A fabricated citation arrives in the same register as a correct one, with the same confidence and the same formatting. The most dangerous output is not the obviously wrong one. It is the plausible one. Unlearning the polish heuristic is the hardest single thing in this training, and it takes deliberate exposure to confident, wrong output to achieve.

What an AI-fluent lawyer can do

The phrase is used loosely, usually to mean somebody who has tried the tools and is enthusiastic. That is not a competence, and it cannot be assessed.

A workable definition is behavioural. An AI-fluent lawyer can do six things.

That definition starts from accountability, not enthusiasm. Current judicial guidance on AI requires generated material to be checked before it is used or relied upon; fluency therefore has to include verification and responsibility for the finished work.

1. Name a task where AI belongs, and one where it does not — with the reason, in terms of what happens when the output is wrong.
2. Write an instruction that produces the same structured result across twenty documents. Not one good result. The same result, twenty times.
3. Diagnose a bad output — distinguish an instruction that was ambiguous from a document that was unusual from a model that fabricated — instead of concluding the technology does not work.
4. Explain what happens to the data. Where it goes, who processes it, what is retained, and why that is or is not acceptable for this matter.
5. Say what the tool cannot do, unprompted, including where it fails.
6. Defend the output to a sceptical partner — which means having checked it against something, and being able to say against what.

Those six are assessable. You can watch somebody attempt each one. That is the difference between a competence and an enthusiasm, and it is the difference between training a firm can point to when a client asks and training that produced a good feeling in the room.

The consequence for how training is run

If the audience already holds the core discipline, the job is not to teach specification from scratch. It is to transfer a discipline they have into a medium where the feedback is different — faster, cheaper, and far less reliable as a signal.

That argues for a particular shape. Work on the participants’ own documents, because the transfer depends on recognising the familiar discipline in an unfamiliar place. Teach failure deliberately, because recognising a bad output matters more than producing a good one and cannot be learned from successes. Keep groups small enough that senior people can be wrong without an audience.

And do not open by explaining what a large language model is. Open by pointing out that they have been doing the hard part for years.