You’re reading criticism, and the first quiet question about any text now is whether a person actually wrote it. Here’s the honest answer for this site — specific enough to hold me to it.

1. The scene — ~250 words, mostly written

I now hire agents from all the major AI labs to do work for me. Last night I asked Anthropic’s Claude Opus 4.7 to research the literature on copy editing and build me a Neovim plugin to copy-edit this essay. It read Strunk, Williams, Pinker, Einsohn, Klinkenborg, King, and Orwell, weighed where they disagreed; built a rule-set that flagged nominalizations, agent-hiding passives, Orwell’s verbal false limbs, and the volunteer phrases Klinkenborg taught a generation of writers to distrust; declined to enforce the zombie rules — split infinitives, terminal prepositions, singular they — that working editors had quietly abandoned decades ago, suppressed any rule it could not back with a citation, and, at my insistence, flagged every em-dash as an AI tell; wrote the code in TypeScript and Lua, tested it against this essay, identified false positives, iterated until it worked. I then asked it to review the code, both as a senior software architect and as the copy-editing authority it had just assembled from those sources, provide feedback, and fix what fell out. By the time I reviewed the code, I had no notes.

2. The wrong question — ~150 words

• Name it plainly: everyone wants a percentage. How much of this did the machine write.
• Two problems, and the second is the real one. There is no stable way to measure it. And even a true answer wouldn’t tell you what you actually want to know.
• What you want to know is whether anyone is answerable for what’s on the page.
• Drop the production question here and don’t pick it back up. Short section. Its whole job is to clear the ground.

3. Where understanding happens — ~250 words

• The claim this page rests on. Understanding is an activity, not a possession — holding a claim well enough to grasp what would refute it, what follows from it, how it coheres with the rest. (Link Understanding; don’t re-argue it here.)
• Writing performs that activity. You find out what you can’t actually say by trying to say it.
• So does dialogue. So does dialogue with a machine about a text — and this is the part most disclosure pages get wrong by pretending it isn’t happening.
• Kleist: thought completes itself in the act of articulating it to someone. His interlocutor is his sister, and it doesn’t matter whether she follows. The presence of someone to say it to is what does the work.
• Collingwood: an answer means nothing except as the answer to a specific question. The machine answers. It never asks. The questions are mine, and the questions are where the understanding lives.
• Land it: the understanding is mine because I am the one holding the claims, the one who can be refuted, the one whose worldview compounds. Its weights do not update, its context ends, it holds no claim as its own. (Link The Work That Remains.)
• This is the section nobody else writes. Don’t hedge it into a defensive paragraph — it’s the strongest thing on the page.

4. What the machine cannot do — ~200 words

• Brandom: to assert is to undertake a commitment and license others to challenge you. Producing a sentence and asserting it are different acts.
• The machine does the first, at any volume, on any subject. It cannot do the second — not because it’s bad at it, but because there’s no one there to be held to it.
• So the line isn’t drawn at production. It’s drawn at commitment. Every sentence here is one I have undertaken.
• Link The Answer Machine for the full development. Borrow the narrow point only — assertion as commitment — and leave aptness-vs-answerability to the essay. Don’t spend the whole argument on a policy page.

5. The rule — ~150 words

• Williamson’s norm, in your own words: I do not assert what I have not checked at the source.
• Why this and not “I wrote the words”: someone can catch you breaking it. Nobody can catch you breaking “the argument is mine.”
• Add the other bright lines you want here — but every one must be catchable. If you can’t describe what a violation would look like, it isn’t a rule, it’s reassurance.
• Test each line by asking: what would a reader have to find to prove I broke this? No answer, cut it.

6. Where the line blurs — ~200 words

• The counter-case. /table §7 is what makes that page credible; this is the equivalent and the page fails without it.
• The real one: a machine surfaced a counter-argument you would not have found, and it changed your position. Whose judgment is that?
• Hold the line — endorsement is an act, and a substantive one if you could have rejected it and know what rejecting it would cost. But stage the difficulty first. Don’t skip to the answer.
• The Kleist tension, if you want it: if the interlocutor’s comprehension is beside the point, why hire the best models? Write into it, not around it. One honest answer is that a better interlocutor produces better questions in you.
• This section converts. It’s the only part a sceptical reader will believe on first pass.

7. The turn — ~100 words

• This page cannot verify itself. It is testimony about testimony.
• And by the site’s own argument there is no detector, no watermark, no test that settles this from the outside — so refuse to offer one. (Link What Cannot Be Faked.)
• What you have instead: a record over time, and a name attached to it.
• Hinge straight into the existing last line.

If any of this ever stops being true, this page changes first. Hold me to it — matt@truetothings.com.