How These Texts Are Written

I write them in conversation with a language model. The material, the arguments and the examples are mine; so is every decision — what goes in, what comes out, in what order and in what tone. A substantial share of the sentences in the final version were written by the model, following outlines I had approved beforehand.

The work looks like this: the conversation is long, and the corrections run mostly in one direction — the model proposes, I reject or rebuild. Models can be useful for checking consistency and catching repetition, but they can just as well propose a logically faulty construction, or attribute to someone a judgement that person never made.

At the end I hand the finished text over for review in a separate session, with no context of how it came about. The model that accompanied the writing remembers everything that was rejected and everything that was agreed, so it also reads what isn’t on the page. A fresh reading is the only way to check whether the text stands on its own. It also happens that only at this stage does the most important thing in the text get named — a thesis covering the whole is visible from a distance, and writing piece by piece leaves no distance.

Everything that ended up staying in the text is my responsibility.