Substrate, not Machine

The decision

In the texts of this project, the word Sprachsubstrat — language substrate — stands where others say AI, LLM, or assistant. This is a decision, not a quirk. It calls for explanation, because the chosen word is unfamiliar and the customary ones are so familiar that their aftertaste scarcely registers.

This note brings the decision into the light.

What a large language model does

A large language model is trained on very many texts and returns, in response to a prompt, a text that is statistically close to the patterns it saw in training. It operates on the surface of language — on what Saussure called the signifiant: the sign, not the signified. It knows word distributions, sentence shapes, idiomatic turns. It does not know what the sentences mean. In Benveniste’s terms: it produces énoncés without the act of énonciation — utterances without the speech-act in which a speaker refers to a world.

That is the technical description. It is sober, not dismissive. Everything that follows rests on it.

What Machine misses

The word machine works as an image — the printing press, the typewriter, the thing in front of one. It works less well when the phenomenon lies within language. A printing press is mechanics. The language model is statistics over writing that was already language. It is closer to a material than to a tool. Whoever addresses it as machine easily hangs a function on it that it does not have.

What AI misses

AI — artificial intelligence — projects human traits onto something non-human. The model has no consciousness, no experience, no standpoint, no world to which it could anchor itself. It imitates the linguistic surface of cognitive acts at great speed. That is striking, and it is not intelligence in the sense the word carried before 2020. Whoever keeps using the word smuggles in a comparability that the matter does not support. The downstream terms — AI collaboration, AI assistance, AI-assisted writing, human-AI writing — extend the comparability and posit two subjects where there is one.

What LLM misses

LLM — large language model — is the correct technical designation, and the term that has taken the place of AI inside the technical speech community. One might expect a model to carry less personification than an intelligence. In the language as it is actually spoken, it does not. The LLM says that… — the LLM answered — the LLM doesn’t know that. The subject-seam has come along. LLM has displaced AI in the developer forums without shedding the anthropomorphizing reflex. What ought to mean a statistical model becomes a speaker in a conversation. Whoever listens knows: what is meant is rarely a matrix of probabilities but rather the thing one talks to.

What Substrate captures

AI and LLM, each in its own way, draw the same line: from what processes language, a speaker is made. The word substrate draws another.

It is the Latin word for underlayer, that which lies beneath. In chemistry, it is the matter an enzyme acts upon. In biology, the ground on which something grows. In linguistics, what lies beneath a language and acts through it. In every one of these uses, the word has the same sense: material that is supplied, without itself being the operation.

Sprachsubstrat carries this sense over to the language model. It supplies material — the statistically likely formulation. What happens with the material is decided by the person at the desk. The act is mediation, the result is machine-mediated writing.

Three things the word secures that AI and LLM do not.

First. It avoids the subject-attribution at the root, not only on the surface of polite phrasing. A substrate supplies, carries, stands by. It wills nothing, says nothing. Responsibility for what stands at the end rests with the one who writes.

Second. It makes a lineage visible. Writing technology — printing press, typewriter, now substrate — is a line in which the substrate is a new but not a first member. Every one of these stations changed how texts come into being without doing away with the one who writes. The substrate is not the exception but the next chapter.

Third. It is abstract enough to mean more than today’s language model. What arrives tomorrow — other architectures, other shapes, other vendors — will fall under the same term, as long as it occupies the same position in the arrangement of writing.

What would be lost

The word is unfamiliar. It is a threshold. Whoever adopts the standard learns it; whoever reads a note should not fail at it. For that reason, Sprachsubstrat is briefly glossed at its first appearance in every note — material that a language model supplies — and then used. Whoever wants to know it more precisely finds the derivation here.

The other words — AI, LLM, language model — are not forbidden. They appear in occasion-blocks where a cited text uses them. They appear in this note. But they do not carry the institutional position. That position is carried by the substrate term, because only that term carries the speech-act of self-declaration cleanly: the author says how much of the language substrate has gone into the work. With AI or LLM, it would be a collaboration ledger, in each case with a different shade. With Substrate, it is what it is: an account of material.

More on the matter: Idea. More on the form of the seal and the axes Wortlaut · Geist · Beleg: there as well. A first application of the position derived here: The Witch Trial.