What Happens to Published Writing When AI Reads for Us?

Something about how I read has changed over the last year or so. I read fewer articles now. If I have a question, I increasingly go straight to AI and keep asking until I understand. If I come across a long piece, I might read a summary, decide I’ve got what I needed, and never return to the original.

Once I noticed this about myself, I started wondering what happens to writing when more people begin reading this way. So I went down a rabbit hole with Claude, trying to work out what the future of writing might look like. And I came away with a few predictions which, taken together, feel less like speculation and more like the shape of what is coming.

Why people read

Start with the basic question: why do we read at all? Look closely and reading seems to do several genuinely different jobs for us. They arrive through the same action—eyes moving over words—which makes them easy to collapse together, but they are not the same.

Call the first row cargo reading. There is something inside the text that you want, and reading is how you carry it out. If someone could hand you the cargo without making you do the reading, you would take the deal. The other rows work differently: their value is created by the reading itself, or by who wrote the thing, or by the fact that other people have read the same words. You can test this against your own behaviour. I only run my summarise-first habit on cargo. I would never ask for a summary of a poem I love, or of the one writer whose voice is the entire reason I’m there. So perhaps my habit is not just laziness; it is a sorting machine. I am classifying writing according to whether its value can be removed from its container.

And AI is the best cargo machine ever built. It reads almost anything, pulls out exactly what you asked for, answers follow-up questions, and shapes the result around what you already know. A published article has to address a thousand possible readers at once; an AI answer is written for one.

Here is the obvious objection: we have seen this before. CliffsNotes. Abstracts. Book-summary apps. Long-form writing survived all of them. Why should this be different? I think there are two reasons, and they arrive together.

First, extraction became nearly free. A summary used to be labour somebody performed once, in the same way for everyone, and it couldn’t answer the next question that occurred to you. Now it is instant, personalised, and conversational. When something expensive becomes free, people don’t merely do it a bit more; they reorganise their habits around it. My embarrassing little ritual is one example. Second—and this took me longer to notice—the new extractor sends almost nothing back. Google answered your question by sending you to a writer’s page (& Meta/X did this indirectly). The writer got a visitor, an ad view, perhaps a subscriber: some evidence that the exchange had happened. Google took a lot, but the loop still closed. AI reads the page so that you don’t have to. By current measurements, AI systems consume on the order of hundreds of pages for every single visitor they return to a source. So the cargo becomes easier to extract at exactly the moment the people producing it stop being paid for the extraction. That is more consequential than a better summary. The loop has been cut.

Why people write

But that is only the reader’s side. Writing is a two-sided activity, and the stranger change may be happening on the other side. Here is a number that reframes the question: the median published author earns about $2000/yr from books. For most people doing it, writing was never much of a job; it was already something else, sustained by other rewards. What rewards? Split the activity in two and it becomes clearer.

There is writing, the private act. Many people write to think—not to record thoughts they have already had, but to have them. You discover what you believe by trying to state it plainly, failing, noticing why you failed, and trying again. This kind of writing needs no reader. AI cannot remove the reason for doing it because the value is in the doing. You can outsource the sentences, perhaps. You cannot outsource the workout.

Then there is publishing, the public act. Publishing runs on evidence that somebody read you: a reply, a subscriber, even a lousy upvote. Study after study finds the same pattern: writers who receive no response quit at higher rates than writers who receive negative responses. Being disliked is survivable. Being unread is not. The audience doesn’t have to be large—fan-fiction writers can continue for years on two loyal readers—but those readers have to be perceptible. The writer needs some sign that another mind was there.

Now put this next to what AI does. Machine reading is invisible. Your work can be consumed more than ever before, flowing through thousands or millions of generated answers, while producing no visible evidence that anybody read it: no visit, no reply, nothing. We have already watched a version of this happen. Stack Overflow was the place where programmers publicly answered strangers’ questions, and it ran on points, reputation, and visible usefulness. After ChatGPT, its activity collapsed to levels last seen in 2009. But Reddit’s programming communities, where people answer one another inside communities that recognise them, did not decline in the same way. Same kind of content, same technological shock. The difference seems to be that one group was producing answers for anonymous traffic, and the traffic disappeared into a chatbot; the other was writing for people who could still see them.

So AI makes two cuts. On the reading side, it takes the cargo. On the writing side, it hides the readers. And between those cuts, one funding route begins to disappear: strangers paying for information through advertisements, clicks, subscriptions, or cover prices. That route paid for much of the writing we have read. It has no defence against either cut, let alone both.

What cannot be carried off?

For a few hundred years, print bundled activities into one object: moving information between strangers and being a human voice. AI is separating them. The information-moving increasingly goes to the machines, but “the voice stays human” is too easy an answer. Plenty of human writing may remain possible while becoming economically invisible, and machines can produce voice-like sentences too. So the more useful question is: what part of this writing cannot be carried off by extraction?

I can find four kinds of protection.

• The first is no protection at all: if the value is cargo, and nothing prevents its extraction, the container is in trouble.
• The second is community. People keep writing when the people they care about can see them doing it; information may escape, but recognition and belonging do not.
• The third is a shared object. Societies need fixed texts everyone can point to—the legal code, the scripture, the classic, the book an entire profession has read.
• The fourth is a particular person. Some writing matters because of who wrote it: their testimony, judgment, reputation, or mind.

Call these moats, with the caveat that a moat only protects a category of writing. It does not guarantee that anyone inside it makes a living.

Some of these conclusions are straightforward, while others become less comfortable when you look at them closely.

Take the person moat. At first glance, it might seem to mean style: a recognisable voice that a machine cannot reproduce. But that cannot be the real protection, because style is precisely the part machines are getting better at reproducing. The harder-to-extract parts are testimony, judgment, and resistance. “I went through this” means something only if someone actually went through it; a machine can generate the sentence, but it cannot have been there. A critic’s judgment matters partly because her reputation is attached to it. If she is wrong, she will have been wrong in public. An AI has no reputation to stake (yet) and pays no price for a bad call.

And the writing that changes you often does so by resisting you—by refusing to say what you wanted to hear. Now notice the tension. AI’s great advantage is that it adapts to you: it knows what you asked, reshapes the explanation, removes the irrelevant parts, and tries again when you object. For cargo, this is ideal. But testimony and judgment require an other, a mind that is not yours and will not automatically bend towards you. A mirror cannot testify, stake anything, or push back. Making it more accurate only makes it a more accurate mirror.

Fiction and poetry are more difficult, and I would rather say this plainly than comfortably. In blind tests, readers already cannot reliably distinguish AI writing from human writing; sometimes they prefer the AI. The penalty often appears only when the machine’s authorship is revealed. Which means the protection around human literature may not be structural at all. It may be a social norm: we value an artwork differently when we know a human made it. Norms can hold for generations. They can also erode. I have a sense that the current norms will fade away and AI-authored writing will become an independent valuable category with it’s own evaluation standards. (The closest precedent may be photography. It took about sixty years to travel from “mechanical trick” to museum art. When photography was finally accepted, painting survived—but photography took the mass market for portraits, leaving painters the prestige end. If machine literature follows that path, human fiction does not disappear; it narrows towards the authors read as themselves, while machines take more of the commodity middle.)

The object moat has a different problem. A world flooded with generated text arguably makes fixed, attributable texts more valuable; we may need things that remain the same because everything else is personalised. But this protects the seat, not the person sitting in it. A machine-written text, once published and collectively adopted, would also be a fixed shared object. The seat remains; the human author is not guaranteed.

Then there is the money. When words become abundant, people stop paying merely to receive words; the payments that remain buy something else.

• Readers pay to sustain a person. Nearly half of paying news subscribers say they are supporting the work rather than simply unlocking it. That is patronage wearing a subscription’s clothes, and AI cannot intercept it because there is no cargo to extract from the act of keeping a person you value at work.
• Readers also pay to belong: to the community, the book club, the comment section that knows their name.
• Institutions pay because some writing is a public good; philanthropy now funds newsrooms at serious scale, and governments increasingly do too.
• And machines pay for raw material—real money, but tending to flow to corporations that own archives rather than to the writers who filled them. Hence one of the stranger facts in this landscape: the largest new market for human writing pays experts $90–150 an hour to produce training material for AI, for words that may never have a single human reader.

Survival can still mean smaller

The moats tell us what might survive. They do not tell us how much of it will remain.

First, reading itself is shrinking. Time spent reading for pleasure fell by more than forty percent over the last twenty years, before AI had anything to do with it, and what remains is concentrating among fewer, heavier readers. Every surviving form of writing is competing inside a shrinking activity. Poetry is the obvious example: almost nothing about a poem can be summarised away, and poetry has been contracting for decades anyway. A moat can protect you from extraction but it cannot make the culture care.

Second, the person moat pays best above a fame line. The surviving discovery channels—recommendations, follower networks, name recognition—route attention through reputation that already exists. Same craft, same theoretical protection, different side of the fame line. The mid-list author and the critic with five hundred subscribers possess the same moat as the stars, and starve anyway.

Third, the fiction question remains genuinely open. If the norm against machine-made art weakens, human fiction may retreat towards the prestige end, as painting did after photography. If the norm holds, human authorship may itself become valuable—the literary equivalent of “handmade.” I don’t think we can know yet.

The writing economy is not disappearing so much as being rebuilt around a different transaction. Readers will pay less often to unlock information and more often to sustain particular people, enter communities, or participate in a shared culture. Institutions will fund the writing they need as a public good. Machines will pay for fresh facts and training material. And somewhere between those systems, writers will still try to build names.

But the old discovery routes are changing. Search increasingly ends with an AI answer instead of a visit. Social platforms still help writing travel, but they keep readers inside their feeds and route attention through follower networks, recommendations, and content that already has momentum. Writers can still be discovered; the path from unknown to known simply becomes narrower and more dependent on platforms they do not control.

That makes the next phase unusually unequal. Established voices become more defensible at the same moment that becoming established becomes harder. Community writers may continue with small audiences because those audiences can see them; equally good writers addressing anonymous traffic may vanish without anyone quite noticing. The question is no longer only whether a kind of writing survives. It is whether we build a way for its writers to be found before they do.