LLMs, a New Stage of Grammatisation

The two intuitions

You asked an AI assistant to draft a briefing note — or a sales proposal, a set of minutes, a delicate message. The text arrived in a few seconds: clean, structured, disconcertingly fluent. And you felt two things at once. First: “this is filler — well turned, but no one thought this text.” Then, more uncomfortable: “I couldn't have done better, and it would have taken me two hours.”

These two intuitions should not be able to coexist. Yet they do coexist, and in the most demanding minds.

Public debate offers to settle the matter. One camp reassures you: these systems are merely “luxury autocomplete,” statistics without thought — sleep easy, nothing important has happened. The other camp marvels or takes alarm: an intelligence is emerging, the machine understands, human writing is living its final decades. The first discourse does not explain the troubling quality of what you have just read; the second explains far too much, and dissolves every nuance. Above all, neither helps you decide what you will do tomorrow morning: whether to delegate or not, what, and on what conditions.

What if the two intuitions were true together? To hold them in a single hand, we need a word the current debate does not possess. The first episode of this series introduced the pharmakon — every technology is at once poison, remedy, and potential scapegoat. Here is the second tool Bernard Stiegler bequeaths to us, more precise, cut exactly for our case: grammatisation.

Discretizing in order to reproduce

The score, or the art of putting the living in a box

The word comes from gramme: the trace, the letter — the discrete unit. To grammatise, for Stiegler, is to transform a continuous, living flow into distinct, manipulable, reproducible elements. His formula holds in five words: discretize in order to reproduce.

The most telling example is not writing, it is the score. A sonata is a flow: a gesture, a breath, a lived time. Musical notation cuts this flow into discrete units — pitches, durations, dynamics — and fixes it on paper. Thanks to this capture, Bach crosses three centuries and is played tonight in Tokyo as in Lyon. But everyone knows what the score does not contain: the touch, the rubato, the presence — everything that makes no two performances alike. The score transmits music precisely because it does not capture all of it.

Three revolutions already behind us

Stiegler reads human history as a succession of these captures, each seizing a flow more intimate than the last. Alphabetic writing grammatises speech: living discourse becomes letters, and this capture — which Plato already distrusted, as we saw in the previous episode — makes law, philosophy, and science possible. In the nineteenth century, recording grammatises the sensible itself: the phonograph captures the voice, the cinematograph movement; Walter Benjamin would analyze what this industrial reproducibility does to art. Then the digital extends the capture to behaviors: clicks, journeys, purchases, hesitations — everything that can be discretized into data.

At each step, the same drama in two acts: an immediate loss (living memory receding before the text, the aura before the copy) and a possible gain (knowledge made transmissible, cumulative, shareable) — never guaranteed, always conditioned on what a culture organizes in return.

What language models capture

Large language models cross a threshold this narrative suddenly makes legible. In them, text is cut into units — tokens, the contemporary grammes — whose patterns of succession the model learns from immense corpora. What is thus grammatised is no longer only spoken speech or observable behavior: it is formulation itself — our collective ways of stringing ideas together, of arguing, of turning a phrase. Centuries of language use, discretized and made replayable on demand.

This is exactly what the most discussed article in the critique of LLMs described, back in 2021. In it, Emily Bender, Timnit Gebru, Angelina McMillan-Major, and Margaret Mitchell called these systems “stochastic parrots”: assemblers of linguistic sequences recombined according to their probabilities of association, “without any reference to meaning.” The phrase launched an endless quarrel: do machines really understand, or do they only pretend?

Reread with Stiegler, the authors' description is rigorously accurate — and the quarrel it set off is badly aimed. A stochastic parrot is an integral grammatisation: the grammes of language without the experience that gave them meaning. But the score does not understand music either — and it carries music across the centuries all the same. The question “does the machine understand?” is probably undecidable; more to the point, it is secondary. The primary question, the one our concrete decisions depend on, bears on us: what becomes of the musicians when the score starts to play itself?

The question that remains when the debate dies down

What history teaches — and what it does not guarantee

The history of grammatisations gives a precious indication: the initial loss was never the last word. Living memory receded before writing, but the school was invented so that each generation would internalize what books externalized. Music theory followed the score. Each time, the gain did not come from the technology alone: it came from the institutions of care — apprenticeships, practices, transmissions — that a society organized around it.

We must, however, lay out the strongest objection to the comfort this history provides: the analogy has a limit. Writing preserved, the score notated — but language models produce. It is the first grammatisation that does not merely record the flow: it generates new ones, indefinitely. The historical precedent therefore guarantees nothing, and anyone who promises you that “it has always turned out fine” extrapolates beyond what history permits. What Stiegler's concept offers is not an assurance: it is a compass — the point to watch remains the same as at the previous stages, and it is not in the machine. It is in what we cease, or continue, to exercise.

The real risk: delegating without internalizing

For the danger proper to this stage can now be stated simply: to externalize formulation while internalizing nothing in return. Drafting all your notes by machine and losing, month after month, the capacity to structure a thought in writing; producing ten times more texts that no one rereads; letting entire teams “write” without anyone still practicing. Stiegler has a name for this dispossession — proletarianisation, the loss of knowledge through its very delegation — and it deserves a whole episode: that will be the next one.

The other path: writing as an author

But externalization can also be accompanied by the opposite movement. Our species has always extended itself through its cognitive prostheses — this is what we call here cognitive hybridisation, and it is not a threat: it is our way of being intelligent. The machine draft violently reworked, contradicted, rewritten until it says what you alone meant to say, can sharpen thought instead of lulling it to sleep — like the pianist who works with another's score and forges his own playing there.

I myself write with language models — this text was prepared, discussed, and contradicted with them, and the site that hosts it owns that transparency. The exercise taught me where the true boundary runs: not between writing alone and writing assisted, but between rereading as an author and rereading as a spectator. The paragraphs I can defend sentence by sentence belong to me, whatever the hand that held the first pencil; those I could not have contradicted never belonged to me.

The score test

From this episode, take away one tool: three questions to ask yourself before delegating a piece of writing — your own or your team's.

1. Could I still do it without? Not in principle: concretely, this week. If the answer becomes no, delegation has turned into dispossession.
2. Am I rereading as an author or as a spectator? The author can defend every sentence, spot the one that betrays his thought; the spectator finds that “it's well written.” It is the same difference as between playing and listening.
3. What am I learning from this delegation? A grammatisation well lived teaches something about one's own knowledge — if only by revealing, by contrast, what the machine cannot do in your place.

At the scale of an organization, the same questions become political: what do we grammatise, for whose benefit, and above all — who organizes the internalization in return? Teams that establish collective rereadings, stretches of unassisted writing, moments where one explains one's text, do exactly what the school did for writing: the institutions of care, at the scale of a team.

And there remains, as always, what the grammes will not capture: thought in search of itself, style in the act of being born, the event of an idea. Improvisation, in short — that part of the playing which appears on no score and which even recording does not fix. It is not a supplement of soul: it is the reserve from which all new writing draws.

The score does not play itself

Three millennia of grammatisation did not kill speech, music, or thought — but it was never automatic: at each stage it took people who went on playing and institutions to demand it. That is the task that falls to us, at the scale of a team, a school, and a family, in the face of the most intimate of captures.

There remains a question this episode has deliberately left open: what exactly do we lose when we delegate — and is that loss reversible? For it Stiegler built his most incisive concept. Next episode: proletarianisation, or what we delegate to machines.