Debates about artificial intelligence and writing tend to begin with authorship. Is a student cheating if she uses a language model to draft a paper? Should a scholar disclose that a model rewrote a paragraph? Is an author less of an author if a machine suggested the words? These are reasonable questions, and institutions will have to answer them. But I increasingly think they focus on the wrong end of the exchange.
My claim in this essay is twofold. First, AI’s capacity to lower the cost of producing language is, on balance, a good thing. It can open the collective intellectual conversation to people whom the old costs of writing kept out. Second, precisely because production is becoming cheap, the pressing question is no longer who writes, but who reads. What is at stake is not just the flow of information, but our participation in a shared human conversation. Writing that no one reads isn’t in a conversation at all. Adapting to AI will require changes in how we write, but it will require even larger changes in how we read.
Language as Collective Cognition
Start with what language is for. Other animals communicate, but human language differs in scale and kind. We describe people who aren’t present, recount events none of us witnessed, formulate rules and abstractions, and, most importantly, pass all of this onward. Language lets an idea move from one mind into another. Writing helps it outlive the mind that produced it.
The result is that human intelligence is cumulative rather than merely individual. The reach of its transmission is astonishing. A story told and retold for generations millennia ago can become a Christopher Nolan summer blockbuster, seen, discussed, and debated by millions, in 2026. Plato began a conversation we are still having. Scientists build on experiments run by strangers. Lawyers argue with judges long dead. No individual holds more than a sliver of what humanity knows; most of it lives outside any single head, in books, archives, legal systems, and institutions.
The value of language, on this view, lies not in the existence of text but in the connections text creates among minds. I read what you wrote; your thought becomes part of mine; I respond; I extend or reject it; someone else carries part of my response onward. The value lies in the loop, and a loop has two ends.
Language does at least two things. It preserves what earlier people thought, building up a collective memory. And it also lets people think with and against one another, which is something closer to collective cognition. Books, archives, databases, and now language models can store or reconstruct a static collective memory. Only our collective cognition requires people to encounter one another’s thought: to interpret it, answer it, revise it, and pass it on. A library is a memory; a living intellectual culture is a conversation. Illiberal societies may have books and libraries—even though they’re not necessarily open to the public. What they lack is candid conversation among political equals—the signal intellectual product of a liberal society. In the age of AI, liberals need to rededicate themselves to the health of those kinds of conversations, and to consider how both their inputs and outputs will change. The rest of this essay is about what happens to that conversational loop, at both ends, when producing language becomes nearly free.
The Unseen Costs of the Old Way
For most of history, producing language was expensive. A book took years of drafting, revision, and review. A serious article took months or more. Before the age of blogs, Substack, and Kindle Direct Publishing, publication required editors, presses, and journals. These production costs forced selection: an author who could write only a few books had to decide which ones were worth writing; a journal with twenty slots a year had to decide which articles deserved them. As a result, products from reputable publishers also carried information. Because producing a substantial work was costly, the existence of the work signaled, however imperfectly, that its author had invested time and judgment. It is the cost of a signal that makes it informative.
These costs, though, were borne unevenly. High production costs create gatekeepers, and gatekeepers were often biased, set in their ways, and self-interested. More subtly, the costs also privileged a particular skill. Access to the written conversation went disproportionately to people who were good at writing, and especially at writing persuasively. That skill correlates with depth of thought, but only loosely. The signal that a reputable publication conveyed was partly a signal of insight and partly a signal of fluency, and the two were never easy to disentangle.
Meanwhile, the costs excluded people who had something valuable to say but struggled to translate it into written form. Some thinkers are slow writers. Some are awkward on the page. Some are working in a second or third language, or have learning differences that make composition laborious, or simply lack the training that elite institutions provide. The historical record of ideas is, in part, a record of who could write well enough to be heard.
AI can change that. A person with a genuine insight and limited facility with prose can now produce a clear, well-organized articulation of that insight. That is potentially a large expansion of the conversation. More people with something to say can now join in, and the conversation is richer for it. Much of the anxiety about AI-assisted writing implicitly treats the old skill barrier as a quality filter. It was partly that. It was also a filter on background, temperament, and training that had little to do with the quality of anyone’s ideas.
None of this is to deny that AI-assisted writing has costs. Critics argue that writing is itself a form of thinking, so that outsourcing the writing hollows out the thought; that reliance on these tools reduces cognitive engagement; and that AI-assisted work tends toward sameness. I don’t contest those concerns. But the existence of costs is not an argument that the costs outweigh the benefits. The old regime had costs, too. We simply stopped noticing them, because they fell mostly on people who never made it into the conversation.
Who Reads?
The trouble is that lowering the cost of writing does nothing to increase the supply of reading.
Human attention does not scale with the production of language. Herbert Simon observed that a wealth of information creates a poverty of attention, and AI is about to test that observation at a scale Simon couldn’t have imagined. If every author can produce more, and new authors can produce for the first time, the volume of language entering the conversation will grow far faster than anyone’s capacity to receive it.
Reading is not simply consumption. It is the receptive half of the loop, the moment when a mind enters into relation with another. An argument that no one encounters doesn’t move from one mind into another. It sits in the archive, contributing to collective memory, perhaps, but not to collective thought. So the question every author now faces is not only “what should I say?” but “who is going to receive this?” If the answer is no one, then producing it, however easy, is not really a contribution. Legal scholars should find this uncomfortably familiar: one study of nearly 250,000 law review articles found that 82 percent had not been cited even once within five years of publication.
The Republic of Letters and the Problem of Scale
Consider the Republic of Letters, the network of scholars who corresponded across Europe in the seventeenth and eighteenth centuries. Its intellectual productivity came not only from the brilliance of its members but from the structure of their exchange. A relatively small number of people wrote in shared venues and a shared style, at a volume their peers could actually absorb. Some people, like the Catholic priest and mathematician Marin Mersenne, acted as hubs and relayed letters among correspondents, including major intellectual figures like Descartes, Pascal, and Hobbes; early journals like Henry Oldenburg’s Philosophical Transactions gave the community a common place to read one another. Participants could plausibly keep up with the work that mattered to them, respond to it, and expect responses in return.
Now imagine that each of those scholars had been able to multiply their output tenfold. One might think the result would have been ten times the intellectual progress. I suspect it could just as easily have destroyed the community. The problem would not simply have been that the best letters were harder to find. It would have been a breakdown of reciprocity. Correspondents would have stopped reading and answering one another. Shared reference points would have failed to emerge. The expectation that a letter would be read and answered, which is what turned private writing into a public conversation, would have collapsed.
The intuition of any individual author in that scenario might be that more production means more chances to be read, more influence, and more engagement. But that intuition doesn’t work for a large aggregate of authors. When everyone produces more against a limited pool of attention, each contribution is less likely to be received, and the shared conversation that made writing worthwhile in the first place begins to dissolve. It is a collective action problem. Individually rational increases in output can produce a collectively worse outcome: a community whose members no longer hear one another.
That, I think, is the real risk of cheap production. An explosion of volume, with no corresponding change in how we receive and sort it, could dissolve the reciprocal attention that turns writing into conversation.
Novelty Is Not the Coin of the Realm
It would be easy to draw a wrong lesson from all this: that the solution to overproduction is to demand novelty, and that anything that has been said before is noise. That lesson badly misunderstands how scholarship works, especially in the humanities and law.
Most scholarship in these fields does not produce new knowledge in the way a laboratory finding does. The questions law professors debate in law review articles, such as the choice between rules and standards, the proper role of courts, or the relationship between law and morality, have often been debated by previous generations, sometimes in strikingly similar terms. That does not make asking these questions unimportant today. We return to some questions again and again precisely because they are important, and because they apply differently as circumstances change. A debate about administrative discretion reads differently in an era of algorithmic decisionmaking than it did in the New Deal. Nolan’s film is not a new Odyssey; its value lies in what the old story means to the people watching it now.
The point of reexamining and republishing these ideas, generation after generation, is to keep them alive in our collective memory. In these fields, relevance to the current discussion, not novelty, is the coin of the realm.
Seen this way, repetition is not waste. A shared intellectual conversation cannot be reduced to a database of unique propositions. Communities keep ideas alive by restating, reinterpreting, contesting, rediscovering, and applying them, and an inherited idea acquires its meaning for each generation through that work.
This sharpens the question of who reads. A genuinely new discovery retains its value even if it sits unread for years; someone can find it later, and it will still be true. A restatement of an enduring idea has no such reserve. Its entire value lies in reaching a present audience and bringing an old idea into a live conversation.
This cuts somewhat in favor of AI-assisted writing. Rearticulating an established idea for a new context is exactly the kind of work AI can help more people do well, including people who understand why an old argument matters now but would have struggled to write it up. The danger lies on the reading side, once AI is doing the filtering. For an information-retrieval system, “this argument has been made before” is sensible evidence of redundancy. For a conversational culture, it is not. A filter built to deduplicate will systematically discount the very work through which ideas stay alive, and a summary reporting that an argument “has been made before” will miss its importance in the current moment.
This is not hypothetical. In late 2025, arXiv, the leading preprint server in computer science, stopped accepting review articles and position papers in that field unless they had already passed peer review. It explained that generative AI had made papers “not introducing new research results” fast and easy to write, and its volunteer moderators could not keep up. Given the volume, the response was understandable—but it may have obscured a lot of conversationally valuable contributions.
Changing How We Write
Part of the response has to come from authors. If the cost of production no longer signals the value of what we say, we need better signals. That might mean writing less, not more, and treating AI’s efficiency as a way to improve the work we produce rather than multiply it. The measure is not whether an idea is new, but whether a piece connects it to a conversation that is actually happening, with readers who are actually in it. And it will likely mean rethinking the forms we write in. Much of our inherited genre structure (the article, the white paper, the monograph) assumed scarcity. It is not obvious those forms remain the right units of contribution when anyone can generate them on demand.
The status incentives surrounding publication make this harder. Academia’s “publish or perish” norms already rewarded volume, and AI makes volume cheaper. An institution that counts outputs will get more outputs. Whether it gets more thought is another matter.
Changing How We Read
The larger adjustment, though, falls on the reception side. If the volume of language is going to grow dramatically, and I think it will, then the practices and institutions through which we find, evaluate, and engage with ideas have to evolve with it.
Some of that filtering will inevitably be done by AI. That is not in itself a problem. Intellectual culture has always depended on mediators: teachers, editors, librarians, reviewers, indexes, and trusted colleagues. No one reads everything, and mediation is how nearly anyone finds nearly anything. The question is not whether we rely on intermediaries, but what kind of intermediaries we are building.
There is a meaningful difference between tools that route readers into a conversation and tools that substitute for it. A system that helps me find the three papers I most need to read, and then sends me to read them, strengthens my participation. A system that summarizes a thousand papers so that I never read any of them does something quite different.
And language models can go well beyond summaries. An LLM can tell me what another scholar argued, lay out the strongest objections, draft my response, and then carry on the discussion with me, playing every other side. The result can be a sophisticated representation of the human conversation, one I engage with fluently while rarely entering the conversation itself. I never have to encounter the other people in it, and they never encounter me. That is a distinctive risk: not that the intermediary fails, but that it becomes good enough that encountering one another stops seeming necessary. A liberal should ask some tough questions about this state of affairs: Who made that LLM representation of the conversation? To what end? What does it include, and what does it leave out? What’s easy to access, and what’s harder? And why does it do that?
This is where the distinction between collective memory and collective cognition matters most. AI can give extraordinary access to the products of collective human thought while quietly withdrawing the user from the conversation that produced them. It can make the user better informed about the collective mind while making them less a part of it. A slightly different technology might bring them into that conversation instead, and which one we build and adopt is a choice.
The reception side will also need renewed human institutions: editors, curators, and communities of peers willing to do the hard work of directing attention to what matters. The Republic of Letters worked because it had shared venues and shared habits of reading. A world of abundant language needs its own equivalents, and it doesn’t yet have them.
Conclusion
The debate over AI and writing has been framed largely as a debate about authorship: who wrote the words, and whether it matters. The more important question is at the other end. AI can make the collective conversation more open than it has ever been, admitting people whose ideas were previously filtered out by the cost and skill of writing.
But the conversation is the point, and a conversation needs listeners as well as speakers. Reading matters because it has long been one of the principal ways we encounter other minds. If we respond to cheap production by producing more in the forms we always have, and if we meet the resulting flood by letting machines read for us, we risk a conversation with more people speaking, but with too few listening. The question is not whether AI will write, or even whether it will read. It will certainly continue to do both. What liberals must ask is whether AI will mediate and enlarge upon the shared human conversation, or whether it will gradually become our preferred substitute for taking part.
Image: Reading by the Sea (1910), by Vittorio Matteo Corcos, via Wikimedia Commons.
