Vocabulary Is Not the Language
Models can master professional vocabulary without yet speaking the language that makes a profession think.
Today's models write law like jurists, medicine like physicians, engineering like engineers. The lexicon is exact, the syntax is the trade's, the tone is indistinguishable. One readily concludes they speak the professions' languages. That is to confuse two things every speaker of a foreign language knows to distinguish: possessing the vocabulary and speaking the language. One can know ten thousand words of a language and remain incapable of saying in it what is needed, to whom it is needed, in the way it is needed. For a language is not a stock of words: it is a system of choices, and the choices that matter are invisible in the words themselves.
What holds for natural languages holds doubly for the languages of trades. A profession is first a language, in the strong sense: a codified way of carving up reality, ranking what matters, saying and withholding. Learning law is not learning Latin words; it is learning to read a situation as a case, to see in a fact a risk, in a sentence a commitment. The vocabulary is acquired in the first year. The language takes ten, because it is not learned from books: it is caught by contact with those who speak it, like every living language.
Translators know this truth better than anyone. Translating never consisted of replacing words with words: dictionaries would have done it long ago. Translating means carrying a thought from one system of carving up the world to another, which is why literal translations are at once exact and wrong. The same holds between common language and a trade language: the same word does not designate the same thing in each. "Fault," "damage," "warranty" have an everyday sense and a legal sense, and the distance between the two is precisely what ten years of practice teach. Whoever possesses the words without the distance produces texts exact and wrong at once.
A profession is first a language, and a language is first a way of thinking.
The grammar words do not show
Every trade language possesses, beneath its vocabulary, a deep grammar the words do not show. It governs what must be said and what must be withheld: the jurist knows that a reservation absent from an opinion is a position taken, that silence on a point is sometimes the heaviest sentence in the document. It governs the weight of terms: "subject to," "on the current state of the file," "it cannot be excluded" are not stylistic prudences but calibrated instruments, whose exact charge every practitioner knows. It governs, finally, the implicit: what a colleague will understand without its being written, what an adversary will read between the lines, what a judge will expect without asking. This grammar appears in no manual, and yet it is what makes a text right or wrong beyond its accuracy.
The decisive point is that this grammar is not only tacit: it is local. Every firm, every legal department, every practice has bent the common language into a dialect of its own. One house never writes "warranty" where another writes it without trembling; one team has its proven formulas, its forbidden turns, its thresholds of caution, forged by twenty years of matters and sometimes by a single bad experience. This dialect is an asset: it encodes, in linguistic form, the organization's private case law, what it learned at its own expense and refuses to relive. To speak the firm's language is to mobilize that memory without even knowing it.
A trade's grammar appears in no manual. Yet it governs what every sentence is worth beyond its accuracy.
The phrasebook speaker
Now measure what the models know. Their mastery of professional vocabulary and syntax is real and spectacular: they have read more legal text than any jurist ever will. But what they read is precisely what gets written, that is, the surface. The deep grammar lives in what does not get written: the corrections a partner makes on a draft, the discussions that precede a formulation, the reasons a sentence was removed before sending. By construction, the general corpus contains final texts and never their second thoughts. The model learned the language as it shows itself, not as it decides itself.
The result is a speaker of a new kind: perfectly fluent and foreign at once. It speaks every trade language the way one speaks with a phrasebook: the sentences are right, the accent is good, and the thought is not there. It will use the exact word without feeling its charge, place a standard reservation where the matter demanded a tailored one, write a perfectly correct sentence that this firm, precisely, would never allow itself. The danger is not the gross error, which shows; it is the prose that sounds right and thinks wrong, the prose that has every property of the language except the one that counts: carrying the judgment of those who speak it.
What makes this speaker formidable is that fluency disarms vigilance. We instinctively judge mastery of a language by ease: whoever speaks without hesitating seems to know what they are saying. This shortcut, reliable between humans, no longer holds here at all, for the model's fluency is total and its relation to the deep grammar, nil. The generated text thus enjoys a presumption of competence that nothing founds: it is read with the confidence granted a native, when it should be read with the wariness reserved for machine translation. Ease, which was a signal, has become a disguise.
The model speaks every trade language like a phrasebook: the sentences are right, the thought is not there.
Where the language lives
If the deep grammar cannot be learned from the general corpus, where does it live? In three places, and their inventory sketches the solution. It lives in the organization's documents, but in dispersed form: a thousand precedents each carrying a fragment of the dialect. It lives in the practitioners, but in tacit form: they exercise it without being able to state it, like every native speaker. It lives, finally, in the corrections, that back-and-forth between drafts and final versions where the firm's language manifests itself in acts: every partner's strike-through is a rule of grammar being stated. None of these three deposits is accessible to a general model, and no future version will change that: it is not a limit of power, it is a matter of access. The firm's language is not on the Internet.
The consequence must be drawn plainly: the trade's language will never come from the model; it must be given to it. Not through ad hoc instructions, which amount to whispering words to a foreigner sentence by sentence, but through formalization: making explicit the house's terminology, positions, thresholds, formulas and prohibitions, and holding them at the disposal of any system that writes in its name. It is demanding work, for it forces one to state what went without saying. But it has a virtue of its own, independent of the machine: an organization that formalizes its language discovers its doctrine, its inconsistencies, and what truly distinguishes it.
The firm's language is not on the Internet. It will never come from the model: it must be given to it.
Making the machines speak your language
Here the question meets architecture. A formalized language has value only if it is actually spoken, that is, if every output of the system passes through it: the terminology enforced, the positions respected, the prohibitions held, the dialect applied to the matter at hand. That is exactly the role of a semantic layer, and that is how MAX is built: the place where the organization's language is made explicit, maintained, and given to the models to speak, whatever the model. The engine brings fluency; the layer brings the language. One commoditizes, the other accumulates.
The dividing line coming between organizations then holds in a simple alternative. Those that formalize their language will make the machines speak it: their assisted productions will carry their voice, their caution, their memory, and every gain in the models will reinforce their singularity. Those that do not will speak, imperceptibly, the language of the machines: an average language, correct, that of the general corpus, the same for everyone. Their documents will be irreproachable and interchangeable, and their dialect, that asset accumulated over twenty years of matters, will fade quietly, not by attack, but by disuse. Languages die that way: not because they are forbidden, but because they cease to be transmitted.
Organizations that formalize their language will make the machines speak it. The others will end up speaking the language of the machines.
