When artificial intelligence meets institutional standards

Generative models are good at language. Counsel are not paid for language.

The current generation of models is remarkably good at producing plausible legal prose. This is precisely the capability institutional counsel should be most careful with, because plausibility is what a good contract dispute is made of. The question is never whether a model can write like a lawyer. It is whether the output survives the standard a lawyer is retained to apply.

A standard is not a tone. It is a set of commitments: that a quoted clause is quoted verbatim, that a cited authority exists and says what it is claimed to say, that a recommendation can be defended in a room full of people whose job is to disagree. A model that is fluent but unverified meets none of these. It meets the appearance of all of them, which is worse.

The design response is not to trust the model more. It is to check it — to have a second, independent pass verify every quotation against the source and every citation against the record before a human ever reads the conclusion. When a citation cannot be verified, it is marked unverified. When it is wrong, it is corrected. The value is not the first draft; it is the audit that follows.

Institutions do not buy language. They buy the reduction of the chance that something confidently stated turns out to be untrue at the worst possible moment. That is a higher bar than fluency, and it is the only bar that matters here.

Used with that discipline, the technology earns its place: it does in minutes the careful reading that would otherwise wait days for a person, and it shows its work.