On September 16, a paper appeared in Nature that does something quietly unsettling to the idea of a paper. A team built a system called Paper2Agent that turns a scientific paper, along with its code and data, into something you can question: an agent that can rerun the paper's own methods on new data and reproduce its results on demand. Nature's own description calls it a "virtual corresponding author." In the team's demonstration, two unrelated genomics papers, each converted into an agent, were set against each other, and an association neither paper had reported on its own came out of the exchange.
It reads, at first, like a UX trick—chat with your PDF, one step past the browser plugins that summarize a page—repairing, without much fanfare, a defect that is roughly 2,400 years old.
Socrates' Second Complaint
The defect goes back to the Phaedrus. Socrates tells the story of an Egyptian king warned against the invention of writing: it would let people rely on external marks instead of their own memory, producing the appearance of wisdom without its substance. That's the complaint most people remember. The second one is sharper, and it's the one this piece is about: a written text, once it leaves its author, can't answer back. Ask it a question and it gives you the same sentence every time, whoever's asking, whatever they meant.
For roughly 2,400 years, this was just true. A book could be reread, cross-referenced, annotated, translated, quoted out of context, disputed for centuries after its author's death—all without ever, itself, responding to any of it. The silence was permanent. What people built on top of it—commentary, footnotes, whole disciplines of interpretation—existed because the text itself couldn't do that work.
Plato's Fix
Plato had a first attempt at a fix, immediately, in the very texts complaining about the problem. Almost everything he wrote is a dialogue: Socrates and an interlocutor, question and objection, false start and correction. It's an odd choice for someone who'd just written that writing can't argue back—unless the dialogue form is itself the workaround, an attempt to smuggle the movement of live argument into an object that has none.
It only gets partway there. Read the Meno and Socrates seems to respond, adjust, concede a point—but Plato wrote every one of those responses in advance. You can't stop him mid-sentence and ask why he accepted a premise; the text performs the same exchange for the millionth reader exactly as it did for the first. The dialogue simulates the shape of a conversation. It doesn't have one.
Everything between Plato and now mostly extends the same workaround rather than solving the underlying problem. The printing press multiplied the silent object without changing what it was. Even the media that feel oral—radio, television, a podcast recorded to sound unscripted—turn out, on inspection, to run on writing underneath: a script, a rundown, a plan sketched out before anyone spoke. Walter Ong called this "secondary orality": voice and immediacy layered back onto a base that never stopped being literate. None of it lets the text answer a question it wasn't asked in advance.
Jeff Jarvis has made the more dramatic version of this argument: that the five centuries print has dominated are a parenthesis, not the natural order of things. Before it, culture circulated orally—conversation, remix, argument with itself. Inside the parenthesis, the object became fixed, authored, permanent. The internet, on his reading, is where that parenthesis starts closing: comments, threads, forwards, wikis, culture arguing with itself again. But that reopening stopped short of the actual defect. A post can be commented on; a wiki page can be edited by a stranger. None of that is the page itself answering the question you asked it—it still just sits there, exactly as before, while the argument happens around it rather than with it.
The Paper That Talks Back
Paper2Agent is a real attempt at that problem. Feed it a paper, and it doesn't just let you ask what a paragraph means—it exposes the paper's own methods as tools you can run, on your own data, checked against what the original reported. Two papers, each given this treatment, can be pointed at each other, which is how the demonstration surfaced a connection neither had drawn on its own. The paper's authors are careful to keep this within limits: results still have to trace back to the original methods and data, credited to the original researchers. The document doesn't disappear; it becomes the thing the agent stays accountable to.
That distinction matters more than the novelty. A summary is still something you read, and still mute once it's written. What's changing is closer to paper becoming capability: a document you could once only extract sentences from is now something you can run. Not everything a paper contains has to be printed out in words for a reader to use it—some of it can just be executed.
Code went through a version of this first, without much of the philosophical baggage attached to a scientific paper. Google's Code Wiki takes a repository and turns it into documentation that regenerates itself with every commit, wired to a chat interface that answers as something that has actually read the code, not summarized a README from memory. A codebase was always technically legible to whoever had the patience to read it; here too, legibility is becoming something closer to conversation.
There's a fitting irony in what one of the most visible tools in this space calls itself. OpenAI's Codex borrows its name from the Latin codex, which first meant the wooden tablets a book was bound between, before it meant the book itself. The support object gave its name to the thing it supported. Now it names the thing that talks back.
The Editor Nobody Elected
The stakes here are easy to understate. In the first months of 2026, an estimated 68% of Google searches ended without a click to any website—up from 60% two years earlier and 49% in 2019. Google's experimental AI Mode is worse for anyone downstream of it: something like 93% of its sessions end without the user ever leaving Google. Publisher referral traffic from search is reportedly down by roughly a third over the past year. None of this is Google drifting into publishing as a side effect of building a better answer engine. It's most of the way there already, and the traffic numbers are the receipt.
A capability that can answer needs someone deciding what it's allowed to say, and that question has an unglamorous, older name: editorial judgment. Google's newly introduced "AI Contribution" program, which pays a site when its content contributes significantly to an answer inside the Gemini app, AI Overviews, or AI Mode, is a first, tentative attempt at pricing that judgment. The mechanics aren't public yet. What's already visible is the shift in what's being paid for: not clicks sent to a page, but a fragment's contribution to an answer nobody will ever visit the page to read. It's a newer verse of an old argument in music: who gets paid when a machine encodes someone else's work instead of copying it outright.
The role being reinvented here is old enough to have a proper history, and it didn't arrive all at once. For most of the print era, printer, publisher, and bookseller were the same trade, often the same shop. As far back as the record goes, it's around the late 17th century that the figure starts to crystallize—Jacob Tonson, for one, bought half the rights to Milton's Paradise Lost in 1683, added the rest by 1690, went on to commission the first properly edited collection of Shakespeare in 1709, and built much of Dryden's career around his own judgment of what deserved backing. He still called himself, on his own title pages, a bookseller, since the vocabulary hadn't caught up with the function yet, but the archetype is already recognizably modern. The full split, publisher as financier and curator on one side, bookseller as retailer of finished objects on the other, doesn't crystallize into two distinct trades until the early 19th century, once print runs and distribution outgrew what a single shop could hold in its head.
Google's version of that role runs the arithmetic in reverse. The publisher's problem was that most of what could be written wasn't worth printing. The problem now is that everything is already printed—indexed, crawled, sitting in some corpus—and the scarce resource is deciding which fragments of it deserve to be recombined into an answer, for this question, for this person, right now. It reads a volume of material no 19th-century publisher could have imagined, and it makes that call fresh every time someone asks it something, instead of once, permanently, at the moment of publication.
What the Silence Was For
Which loops back to Plato more directly than any of this looks at first. He wrote dialogues because dialogue was the only literary form he had for smuggling response into an object that had none. Paper2Agent, and whatever Google's editorial layer becomes, are both attempts at the same smuggling operation, just with better tools: not writing that imitates the shape of an answer, but writing that can produce one. Twenty-four hundred years of reading the Meno as a performance of a conversation that can't actually happen, and something now exists that might let it.
Socrates called his own method maieutic—Greek for the art of drawing out what's already there: he claimed never to hand over a conclusion, only to ask the questions that let someone deliver one they already carried. A tool that just supplies answers on demand isn't practicing that at all—if anything, it's closer to what Socrates was arguing against in the first place. Pushed instead to interrogate a half-formed argument, find the weak premise, produce the counterexample nobody had thought of, the same tool starts doing something that actually resembles maieutic work, which a silent book, however well written, could never quite manage on its own.
But the silence was doing more than just failing to answer. A book that couldn't respond forced the reader to do the responding: hold two passages against each other, go find the counterargument yourself, sit with a contradiction instead of asking your way out of it.
That friction was the exercise, not a leftover flaw from before writing worked properly. It's the same argument playing out a level down, in what happens once shipping software gets too easy: removing friction from a process doesn't automatically improve it, because some of what the friction was doing was perhaps the point.
