Talking to yourself

Every writer knows the hour, and none of them describe it the same way. You sit down, and the thing in your head is clear enough, and the page will not take it. The words arrive in the wrong order, or in the wrong register, or they arrive plainly and turn out to be the wrong idea. You cross the sentence out. You write it again. The distance between what you meant and what you can get onto the page is the oldest problem in writing, and it does not care how good you are.

What is newer is the arrangement around that hour. For about a century and a half, writing in this part of the world was taught as one skill with one correct form. In 1783, Noah Webster set out to make every child in the new country "speak and spell the same way." A century later, Austin Palmer sold the schools a method of handwriting built on arm movement and drills, and its stated object was to "produce a standardized, easily read script." The school essay arrived with the same logic, and so did the red pen, the workshop, the submission queue, and the publisher's list. The industrial arrangement did what industrial arrangements do: it found one best way, taught it to everyone, and charged admission at the door.

Anyone whose hands do not hold a pen that way, or whose eyes or attention do not run in that line, or who never had the schooling, was told early and clearly that writing was not theirs. That was never a fact about them. It was a fact about the arrangement, and it is the reason this appendix exists.

So, plainly

This book was written with an AI assistant. The conversations it quotes come from the Future Herd podcast, and their transcripts came through a machine that turns recorded speech into text. The assistant read them, assembled the research, drafted prose, ran the checks that count words and hunt banned phrases, and kept the record. I set the argument, chose the voices and refused some of them, wrote the standard the drafts had to meet, and rewrote or threw out most of what came back. The sentences are mine to answer for. That is the arrangement, and it is not a confession.

Two properties of the assistant matter to everything below. The first is that it decided nothing. It has no opinion about farming, no stake in this book, and no idea what the book is for unless I say so. The second is that it cannot put a person into the book who is not already in the record, and I cannot either, because every named person here was checked against a recording, a transcript, or a saved page. It can still be wrong in ways that look like success. It handed a sentence to the wrong speaker once. It described a guest with a title he had already left. Both errors were caught by the check, and both are written down, which is the only defence available against a machine that stays fluent while it is wrong.

What the objection is really about

The objection to all of this is not stupid, and most of it is not about writing. It is about a flood. In February 2023, the science fiction magazine Clarkesworld temporarily closed its submissions for the first time in more than a decade, after a wave of machine-made stories arrived hoping to be published with none of the work done. Its editor, Neil Clarke, found a phrase for the motive that is hard to improve on: "an honest interest in being published, but not in having to do the actual work." He had measured the tide, and the measurement is blunt. The share of spam submissions resulting in bans, he wrote, had "hit 38% this month."

The labour objection is separate and heavier. The Writers Guild's 2023 contract, settled after a strike that stopped American television, drew the authorship line in a sentence: neither traditional nor generative AI "is a writer," so nothing it produces counts as literary material. A writer may use one, on the writer's own choice and with the company's consent, and no company may require it. The law has been circling the same conclusion. In January 2025, the U.S. Copyright Office found that the output of a generative system is protected "only where a human author has determined sufficient expressive elements," that supplying prompts is not enough, and that using a machine to assist in making a work "does not bar copyrightability."

So the line has mostly been drawn, and it has been drawn where I would draw it: the person is the author, and the tool is a tool. What is left is not a question about machines. It is a question about who gets to write. Take the people with the strongest reason to resent these tools, the literary translators, whose work is the easiest to imitate and the most exposed. In a survey of nearly eight hundred authors, around a fifth said they had used generative AI in their work. Among translators the figure was a third, the highest of any group in the survey. The people nearest the threat were also the ones using it, which suggests the interesting behaviour is neither loyalty nor betrayal. It is judgement, made case by case, by people who need to work.

The worst version of this appendix's own case was also made in public, and it is worth naming so as not to repeat it. In September 2024, the organization behind National Novel Writing Month published a statement observing that "the categorical condemnation of artificial intelligence has classist and ableist undertones." Its members heard that as being told that their objection was bigotry. Board members resigned, the community split, and the organization closed the following year. The claim was not false. It came from an institution defending a tool, addressed to people whose livelihoods were the argument, and an argument about access that lands as an accusation stops being about the writing. The case for assistance is not a case for the machine. It is a case for the person, and it is much older than the chatbots.

The tool that was built for someone

Word prediction is the clearest example, and it is old. It exists because a person who cannot easily type should not have to pay full price for every word: the program guesses the rest of the word from the letters you have, so a sentence costs fewer keystrokes, fewer movements, and fewer attempts at a spelling that keeps going wrong. Teachers and parents describe what that is worth better than any policy could. One mother, writing about a word prediction program and her son, a keen reader in primary school who was falling behind on writing: "the technology has made such a difference to allow him to take ownership of his own learning." Nobody in that house was arguing about the future of literature. They were trying to let a child write.

Captions have the same shape of history, and the same habit of ending up everywhere. On 5 August 1972, a cooking program made in Boston became the first national broadcast in the United States that deaf viewers could follow, which sounds like a small thing until you read what it actually was: the first time that deaf and hard-of-hearing Americans "could enjoy the audio portion of a national television program." Captioning is now a default on every screen, and its heaviest users are not the people it was built for. They are anyone on a loud train.

Curb cuts tell it best. One night in the early 1970s, a group of wheelchair users in Berkeley poured cement into the form of a crude ramp and rolled off into the night. "The police threatened to arrest us," one of them, Michael Pachovas, remembered later. "But they didn't." Berkeley installed its first official curb cut in 1972, the idea spread, and it ended up in law. Then something happened that nobody had promised: the wall came down and everybody used the gap. Parents pushing strollers went straight for it. So did workers pushing carts, travellers wheeling luggage, runners, and skateboarders. A study at a Florida shopping mall found that nine out of ten pedestrians who did not need a curb cut went out of their way to use one. Angela Glover Blackwell gave the pattern a name, and the reason to keep it in mind is the direction of the traffic. The device was designed for the people it was hardest to include, and it is used by everyone.

The newest assistance follows the same line, which is the finding that surprised me most. The American Foundation for the Blind surveyed more than a thousand AI users with disabilities and more than six hundred without. Disabled users were far likelier to use the functions that carry information from one form into another: "Almost half (44%) of disabled participants used AI for visual descriptions compared to just 26% of nondisabled participants," and 43 per cent used it for captions. Access is not a courtesy extended to a few people at the end of the discussion. It is where this technology is already most in use, by the people who saw a use for it first. The same study carries the other half of the warning, and it belongs here rather than in a footnote: its disabled respondents were almost three times as likely to report a health insurance denial in the previous two years. That is the same machinery deciding something about a person instead of deciding something with them, and it is the difference between an assistive tool and a gate.

Access runs in two directions, and the second one is the reader's. Getting the words out of the writer is half of it. The other half is getting the meaning into the reader: plain language, short sentences, every technical term explained the first time it appears, and prose that survives being read aloud, because a book that only works on a printed page excludes the people who listen. This one is written for both, and neither choice costs the writing anything. The improvements that arrive first for the readers nobody designs for turn out to be improvements for everyone.

A calculator that predicts

Now the machine itself, plainly, because most of this argument is lost to vagueness. A language model is a program that has read an enormous amount of text and learned which words tend to follow which other words. Ask it what comes next and it estimates, from everything it has seen, what a passage like this one usually says next. That is the whole mechanism, and it is worth taking at face value: it is a calculator that predicts. It has no experience of a farm and no memory of the season it is describing. It does not know what a harvest weighs, or what a person does the morning after a bad year. It is not alive and it does not think, and saying so is not an insult to it. It is a description.

And then the interesting part. When a person writes with one of these tools, the conversation is not between a writer and a machine. You put a sentence in, and what comes back is a version of what you meant, or of what you nearly meant, in language gathered from other people. So you are talking to yourself, with a library listening. The private half of that is the ordinary strange thing about writing: you do not know what you think until you have said it and heard it back. The public half is the internet, because the model's answers are assembled from an enormous record of what everyone who wrote before you had to say about the same questions. Your own unconscious, and the collective one, arriving in a single voice that belongs to nobody. That is why an argument about the machine misses the point. There is no author in there to argue with.

What to judge instead

The question worth asking about any piece of writing was never about the instrument. It is whether the thing is true, whether it is useful, and whether it tells the reader something they did not have. A sentence is not better or worse because of the tool that helped carry it, any more than a barn is less of a barn because the lumber was cut by a machine.

Two questions I cannot settle and will not pretend to. The first is whether a practice of disclosure can hold when almost nobody can verify it. Saying a thing, as this appendix demonstrates, is the easy half. The second is what these tools do to the writing of people who were already being told that writing was not theirs. If an assistant gets the sentence out of them, it is a door. If it gets the sentence out of them and they never learn to make one of their own, there is a turnstile in the door, and the difference will not show up in a book like this one for years.

Two things worth doing this week. When you meet a piece of writing whose origins are in question, ask what the tool did. Most of the public argument has skipped that question to argue about the keyboard. And if you write with one of these tools, say so in one sentence the next time you publish something. A practice of disclosure is worth more than a purity test, and it is the only norm here that any of us can actually build.

The Situation takes an hour and needs nothing. Write one paragraph about something you know first hand, and ask a writing assistant to continue it. Keep the exchange. Then write down the first suggestion you refused, and why you refused it. The first half shows you what the machine does. The second half shows you what you do, which is the part no tool supplies: a person with something to say, deciding when it has been said.

Less about the machine, then, and more about the ideas we are trying to hand each other. There has never been a shortage of people with something to say. There has only ever been a shortage of ways to get it out, and that shortage just got smaller.