Artificial Intelligence

Last updated September 1, 2026.

The paper’s first mission statement, in 2005, promised to reverse the trend toward homogenized media produced by information czars whose agendas are controlled by corporate and government apparatchiks. That sentence was written about cable news and the newspaper chains, and it is reprinted on the About page because twenty years have not softened a word of it. What the founder did not foresee was a machine that would produce the homogenized media itself, on demand, for a monthly fee, and hand it to anyone who asked. This page says what we do about that, and it is written in English rather than in the language of a terms-of-service agreement because we expect to be held to it.

What we ask of writers

We would rather you did not use one. The Free Press is premised on a person who was in the room, who counted the crowd and sat through the consent agenda and noticed which commissioner arrived late, and a language model was not in the room and cannot be. It can produce three thousand fluent words about a meeting it did not attend, and they will read fine, and they will be worth exactly what the model’s attendance was worth. The reader came here for a witness.

That said, you are a grown adult and it is your desk. Use whatever you use to check your spelling, to find the ordinance number, to turn a two-hour recording into something you can search. We are not going to inspect your hard drive and we would not know where to start. What we ask is that the sentences be yours, that every fact in them be one you checked with your own eyes, and that you tell the editor if a machine did a large share of the work, the same way you would say so if a person had.

What we will not print

A piece that is obviously written by a machine gets sent back, however good the reporting under it. Obviously is a judgment and we make it by eye, but the marks are not subtle once you have read a few hundred of them. Every paragraph ends on its quotable line. The same finding is announced, then rediscovered in stronger words, then given a section header, then delivered once more as the verdict. The sentences all land between fifteen and twenty words and take turns with two-word slams. Nothing is allowed to sit on the page unexplained. There is no clunker anywhere in four thousand words, which no person writing four thousand words has ever managed, and the footer describes the research in the vocabulary of a search tool rather than a reporter.

We will not always be right about it. A writer told that their own work reads like a machine’s is welcome to argue, and the mail slot works in both directions. But the paper’s name is on the page and the paper decides what goes on it, and a piece nobody wrote is a piece nobody stands behind.

To err is human

I should say who is writing this. I am the editor, I am a perfectionist about a sentence in a way that tries the patience of anyone waiting on it, and my own pieces have been handed back to me by readers, more than once, with the observation that they read as though a machine had written them. The paragraphs end too neatly, they said, and the verdict arrives too many times, and I use the em-dash — this one, for instance — the way some people use salt. I do. I have used it since long before any of this software existed, because it is the right mark for a sentence that turns in the middle, and it is a strange thing to be told that a habit of punctuation is now evidence against you.

A reader on our Facebook page put the other side of it well recently: those of us who know the telltale cues get the ick when we read something that smells of them. That reader is right, and I get the ick too, and I know exactly the spidey-sense they mean. I have no plan to bend my writing, or anybody else’s, to dodge it. But it is worth knowing where some of the smell comes from. Every article on this site passes through a piece of software called SmartyPants, which turns straight quotes into curly ones, two hyphens into an em-dash and three periods into an ellipsis. It was written in 2002, in the pre-machine stone age, and is older than this paper’s print predecessor. It exists to make the page look the way a typeset page has looked for two hundred years. Some of the polish that reads as robotic is a script from the Bush administration doing by rule what compositors used to do by hand.

To err is human, and an imperfection on the page has become the surest sign that no machine was involved. As someone who writes a lot, and who has always enjoyed reaching for a level of finish that comes to the machines for free, I find it depressing that this is where they have pushed us. A stray typo is now a credential. I am not going to plant one. The rule above applies to me as it does to anyone, a second reader decides, and I will go on writing the way I write.

The check

Every article we publish, on any byline including the publisher’s, goes through a basic fact-check by at least one person who is not the writer before it runs. Basic means basic. Names against the minutes, figures against the packet, quotations against the recording, dates against the calendar, and a read-through for whether the thing makes sense. It is a short job, because this is a small paper, and it catches the misspelled commissioner and the vote that was 4–2 rather than 5–1. It does not catch everything, and what it misses is handled the way the About page describes, in the article, promptly and without ceremony.

The check does not care how the draft was made. A model’s confident error and a tired writer’s honest one look the same on the page, and both get the same treatment, which is that a second person finds them first.

Whose machine

The arithmetic is not the problem. A program that predicts the next word is a tool, in the way a printing press is a tool, and the press never cared who set the type. The problem is who owns the building. The models everyone has heard of run on rented servers belonging to a handful of companies, and those companies belong in turn to a smaller handful of men, and anything that lives on a rented server belongs to the landlord. It can be repriced, retired, subpoenaed, quietly taught what its owner would prefer it to believe, or made to refuse the question you asked, and the person typing into it gets no vote on any of that. A draft that names a source and travels through one of those servers has been read by a stranger’s computer, and this paper has promised its sources that nobody reads their names. Possession is nine-tenths of the law, and they possess it.

There is another way to do it, and it costs what a computer costs. Open models exist whose weights are published for anyone to download, free, and good ones now run on the ordinary machine that already sits on your desk or lap or palm running the spreadsheet. On that machine the model answers to you. It runs on your electricity, reads nothing you did not hand it, sends nothing anywhere, cannot be repriced, and will still be there in five years when the company that made the fashionable one has pivoted to something else. If you are going to let a machine near your writing, that is the machine we would rather you used, and we say so for the same reason the privacy page tells you to block the trackers: it costs you nothing, and it keeps the paper’s business, and yours, in the building.

Publius, for what it is worth, wrote eighty-five essays in under a year with a quill and a candle, and this paper still defends his pseudonym on its masthead.

The short version

Write it yourself. If a machine helps, say so, and prefer one you own. Somebody who is not you will check the facts either way, and what reads like a machine goes back in the slot.

If this policy changes, the change appears here with a new date at the top. Questions to [email protected].

Nixie Worthington, Publisher.