AI Vocabulary

Context: why does a good model produce a bad document?

Context is everything one must know about a matter to produce the right document: the client's positions, what was negotiated, what was set aside and why. An excellent model without it produces an irreproachable and generic document — that is to say, unusable as it stands.

What the word covers

Not the documents, which are accessible, but what surrounds them. That this client systematically refuses exclusivity clauses. That the liability cap was obtained in exchange for extended warranties. That ownership of developments was settled two years ago and must not be reopened.

None of that appears in a contract. It lives in the memory of those who negotiated, and passes orally when someone picks the matter up.

Why the model can do nothing about it

A model learned on general corpora. It knows the law, drafting conventions, the structure of an instrument. It does not know your client, and no increase in its power will change that — it is not a question of capability but of access.

This is why the gap between a middling and an excellent model is small on a contextualised task and considerable on a generic one. On the first, what the weaker one lacks the better one lacks too.

What distinguishes context from documents

A confusion is worth clearing up, because it leads to overestimating what access to files provides.

Giving a system access to every document in a matter gives it access to what was signed, not to what was decided. The difference is that between the outcome of a negotiation and its course: the contract says what was agreed, it says neither what was asked for and refused, nor what was conceded in exchange.

It is precisely that part which constitutes context, and it exists in no document.

The symptom that reveals it

The document produced is well written, legally correct, and it does not do. It addresses a point this client does not want raised, it omits a hard-won qualification, it reopens a settled trade-off.

The reviewer then spends longer correcting than they would have spent drafting, because they must first identify what is wrong in a text with no visible defect.

What it does not solve

Supplying the context at each request is possible and costly. It is what those who obtain good results do: they spend an hour explaining the matter before asking anything.

That cost does not amortise. Next time it must be repeated identically, because nothing was kept. This is why most practitioners give up after two attempts and return to short tasks, where context fits inside the question.

And retained context does not replace judgement. Knowing that this client refuses exclusivity does not say whether it should be refused this time.

Why it matters to a lawyer

Because it explains the most frequent disappointment: the tool is excellent in demonstration and disappointing on real matters. The demonstration covers an isolated task, where context fits in the statement; a matter never has that property.

The question to ask is therefore: what I explain to it today, will it still know next week? If the answer is no, every use will start from nothing.