To Publish or Not to Publish

In the decades before 1665, a natural philosopher who discovered something faced an awkward problem. Announce it, and he might be believed, or robbed. Stay silent, and someone else might find it and take the credit. Galileo had met the difficulty by publishing some of his discoveries as scrambled Latin anagrams, so that he could reveal the solution later and claim priority without handing rivals the finding in the meantime.¹ Discovery existed, but it had nowhere reliable to live. There was no dated, public, attributed record that could say: this person knew this thing at this moment.

In 1665, the Royal Society helped solve the problem by building a new kind of container. Its secretary, Henry Oldenburg, began printing Philosophical Transactions, a periodical that took private findings and gave them a fixed public form: a date, an author, a claim, and a place in a record that others could read, cite, and contest.² It became one of the oldest scientific journals still published today, and it did something more consequential than circulate results. It turned thought into an object. Once an idea was published, it existed in the social world in a way it had not before. It had priority, a claimant, and an address in collective memory.

This is worth seeing clearly, because AI is now doing something to that object.

The Old Container

Publication was never merely a way to spread knowledge. It was a social technology that solved several problems at once: attribution, priority, memory, contestation, and recognition. It let a society decide which private thoughts deserved public standing, and it let individuals convert what was in their heads into something the world would remember and reward. For three and a half centuries, the book, the article, the thesis, and the degree have been the containers in which that conversion happened.

Every such container has held two different things at the same time, and their bundling is the key to everything that follows.

The first is knowledge value. A book can teach you something; an article can report a result or move a field; a textbook organizes a domain. This is the container as a device for carrying understanding from one mind to many.

The second is signal value. The same object also tells the world something about its author. A peer-reviewed article signals that someone passed a field's certification. A book from a serious press signals that editors and reviewers judged the work worth attention. A degree signals passage through a filter. A citation record signals recognition. None of this is simply about the knowledge inside the object. It is about the standing the object confers on the person who produced it.

For most of the container's history these two values could not be pulled apart, because you could not get one without the other. To earn the signal, you had to produce the knowledge; and the ordinary way to distribute knowledge was through an object that also carried the signal. The book was a container of understanding and a status object in one. The article was a contribution and a membership token in one. That fusion is what made publication so powerful. The reader gained knowledge while the author gained recognition, and neither was easy to fake without the other.

The whole apparatus of modern intellectual life was built on this fusion, and priority disputes are the clearest evidence of how much the signal mattered. Newton and Leibniz, and their followers after them, fought for decades over who had invented the calculus, precisely because the timestamp was the currency. Newton had the mathematics early and published late; Leibniz published first and secured a claim that Newton's partisans never forgave.³ The knowledge itself eventually stood on its own. The signal did not.

AI Opens the Container

AI breaks the fusion, and it breaks it asymmetrically.

A base AI model can be understood as a kind of fixed-point machine built from the accumulated archive of human thought. It condenses books, articles, code, manuals, public discourse, and institutional language into a usable projection of what the knowledge system has already digested. When asked about a field, it can quickly return the standard story, the major concepts, the central debates, the conventional gaps, and the likely shape of the conversation. What once required long reading, scanning, and synthesis can now be approximated in compressed form.

That does not make AI an original discoverer. But it does change the time structure of knowledge. It raises the floor of background understanding across competitors. The old advantage of having read and synthesized more historical material becomes less scarce when many people can reach a usable map of the field quickly. The scarce edge moves closer to the present: latest data points, unpublished observations, private notes, proprietary evidence, live retrieval, tacit judgment, and what has not yet been absorbed into the machine's fixed point.

It helps to name what kind of thing this condensation is, because it is itself a container of knowledge, only of a new kind. The old container froze one argument at one moment and handed it to you closed; the model melts millions of arguments into a single fluid surface that reshapes itself to whatever it is asked. That is its power and its particular defect. A condensation returns the settled center of what has been written and quietly drops the rest: the dissent that never won consensus, the craft no one wrote down, the trail of who found what and why it was once surprising. It hands you the field's center of gravity and often hides the field's edges. The opened container does not so much make the old one obsolete as invert its value, and the smoother the condensation grows, the more the weight shifts to what it cannot hold.

The consequence for the publication container is direct. What falls, at the margin, is not the object's knowledge value but its value as an entry point, because the object is no longer the only way in. You do not have to read the book to get much of what the book knows. You can ask for it, interrogate it, personalize it, translate it, and receive it in a form the author never anticipated. As an entry point to knowledge, the container is being opened.

The signal value moves the other way, into crisis. When polished, article-shaped text becomes cheap to produce, the existence of a polished, article-shaped object certifies far less than it used to. The object was always a proxy. It stood for the human effort, training, and judgment required to make it, and it worked as a signal because that effort was hard to counterfeit. Once the effort becomes cheaper, the proxy comes loose from the thing it was standing for. The field is left asking questions the finished object used to answer on its own: who actually understood this, who generated the idea, who can defend it, what in the work belongs to the author, and what belongs to the machine?

So the two values that the container fused for three centuries come apart. Knowledge value is liberated and diffused. Signal value is thrown into question and has to be re-secured by other means. That split, far more than the bare fact that AI can write, is what reorganizes the economics of publishing.

The Disclosure Dilemma

Once the values separate, publishing stops being an obvious good and becomes a strategic decision.

Under the old bargain, the decision was easier. Publish, or your knowledge did not fully count. Silence meant no priority, no citation, no institutional recognition, and the standing danger that someone else would publish first and take the record. Darwin lived that fear in its purest form. He had worked out natural selection by the early 1840s and then sat on it for the better part of two decades, refining and hesitating, unwilling to release it, until in 1858 a letter arrived from Alfred Russel Wallace setting out substantially the same theory. The threat of losing priority did in a few weeks what twenty years of conviction had not. A joint announcement was arranged at once, and On the Origin of Species was in print the next year.⁴ The container demanded disclosure, and the price of silence was very nearly the loss of the claim.

AI keeps the old reward and adds a new cost. Publish, and private thought still becomes public reputation. But publish, and it also becomes machine-digestible material: something that can be summarized, imitated, recombined, and absorbed at a speed the old human reading economy never had. The author still gains priority and standing. The author also releases the insight into an environment that can strip its novelty quickly. Publication has become an act of conversion in two directions at once. It turns private thought into public capital, and into public input.

This is why the sharpest knowledge holders are converging on a layered rule: publish the anchor, protect the engine.

The anchor is the public claim that establishes authorship and standing: the paper, the book, the report, the patent, the lecture, the visible thesis. The engine is the machinery underneath: the data, the workflow, the current retrieval, the tacit judgment, the private notes, the live interpretation. You publish the thesis and hold the dataset. You publish the framework and keep the operating procedure. You publish the mature result and delay the frontier observation. The general form is simple: disclose what is worth more for being known, and protect what is worth more for being scarce. Private quantitative investment funds are the limiting case, publishing almost nothing while the world reconstructs their standing from returns alone.

One caution the neat rule hides is that the anchor can leak the engine. A capable model can often infer more than the author meant to disclose, reconstructing a method from its reported results, approximating a dataset from its summary statistics, or interpolating the judgment behind a framework from the framework itself. A published claim is no longer an opaque surface laid over the machinery beneath it; it is a set of clues that a system can run backward. So the separation is leakier than it looks, and the real question is not only what to withhold, but how much of the hidden work any disclosure drags into the open along with it.

That is one strategy, and stated this way it can sound obviously right. It is not.

The opposite strategy is at least as strong, and in a world of cheap copying it may sometimes be stronger. If an idea will be digested anyway, the winning move can be to publish it first, fast, and loudly, planting the flag before a rival or the machine gets there and converting the insight into reputation while it is still yours to claim. Openness compounds. A disclosed standard becomes the one everyone builds on. A disclosed method attracts citations, collaborators, students, criticism, and extension that secrecy forecloses. An open body of work can throw off reputational returns for years, while a hoarded one earns only until someone else finds the same thing. The case for disclosure is not naive generosity. It is that in a fast-copying world, being first and known can dominate being sole and hidden.

Which strategy wins depends on where the value actually sits. The anchor-and-engine split is the synthesis of the two. Knowledge whose value is recognition should be published early and openly, because recognition exists only once it is public and accrues most strongly to whoever is first. Knowledge whose value is exclusivity, the live dataset, the proprietary method, the not-yet-obvious observation, should be timed and gated, because its worth may evaporate on release. The error is to treat all knowledge as one kind. The discipline the AI era demands is the habit of asking, of each piece, whether it earns more by being seen or by being kept.

Academia is the one place that cannot fully play this game, and its predicament is the sharpest form of the whole problem. Scholarly legitimacy rests on inspectability. A science in which everyone claims private evidence that cannot be examined is not a science, so the academic bargain requires disclosure, replication, and public contest by design. Yet the forces that raise the value of withheld knowledge everywhere else press on the university too. As AI digests published work ever faster, the relative worth of what remains unpublished rises, and the scholar who discloses everything hands the machine and the field a fully digestible artifact in exchange for a citation. The institution rewards publication; the technology rewards control of the undigested. Academics will feel that contradiction earlier and more acutely than anyone, and they have the least freedom to resolve it by withholding.

The New Signal

None of this means the container disappears. It means its job changes.

Society still needs anchors: fixed points with a date, an author, and a structure of evidence that can be cited and contested. It needs them precisely because machine-generated conversation is fluid, unattributed, and detached from lineage. What falls is the container's monopoly on access. What rises is its function as certification. The book matters less as the only door into its contents and more as a dated, authenticated claim about who knew what, when, and under what discipline of evidence.

And the signal, having come loose from the polished object, has to reattach somewhere. It is migrating from the artifact to the process behind the artifact. When a finished text proves little, the questions move to what a text cannot easily fake: verified authorship, a replication package, a live defense, demonstrated judgment under questioning, a record of the thinking that produced the claim.⁵ The old container certified by its mere existence. The new one will have to certify hidden formation, the part of knowledge that lives in a person and can still be tested when the machine is taken away.

But process is expensive to inspect in a way the finished artifact was not. A journal could certify finished artifacts at scale; watching someone think, checking whether a result truly reproduces, or questioning an author in real time does not scale in the same way. The likely result is concentration. A few figures who can afford to perform their process in public absorb most of the signal, while everyone else struggles to have their process seen at all. The old container gave a modest signal to many. The new one may give a strong signal to a few and almost none to the rest.

The Royal Society built its container to answer one question: what have you published? For three and a half centuries that question organized memory, status, and the relationship between private thought and public standing. AI does not abolish it, but it adds two questions that cut deeper.

What do you know that has not yet been digested?

And can you prove you are the one who knew it?

To publish or not to publish was once a choice between silence and standing. It has become a decision about when private thought is worth more as public capital than as private advantage, made in a world where every frozen thought can now be thawed, absorbed, and sent back out in someone else's voice.

Notes

1. Galileo encoded discoveries such as Saturn's unusual appearance and the phases of Venus as Latin anagrams in 1610, so that he could establish priority while withholding the finding until he chose to reveal it. On secrecy and priority in his practice, see Mario Biagioli, Galileo's Instruments of Credit: Telescopes, Images, Secrecy (Chicago: University of Chicago Press, 2006).
2. Henry Oldenburg, ed., Philosophical Transactions, vol. 1 (London: Royal Society, 1665). The first issue appeared on 6 March 1665, making it one of the earliest and longest-running scientific journals.
3. On the priority war over the calculus, see A. Rupert Hall, Philosophers at War: The Quarrel between Newton and Leibniz (Cambridge: Cambridge University Press, 1980). Leibniz published his method in 1684; Newton, who had developed his own years earlier, published later, and their partisans disputed precedence for decades.
4. Charles Darwin and Alfred Russel Wallace, "On the Tendency of Species to form Varieties; and on the Perpetuation of Varieties and Species by Natural Means of Selection," Journal of the Proceedings of the Linnean Society of London, Zoology 3 (1858): 45-62, read on 1 July 1858. Darwin's On the Origin of Species followed in 1859.
5. In January 2023, Nature and Science both ruled that large language models cannot be listed as authors and that any use must be disclosed, an early institutional attempt to re-secure the authorship signal. See the editorial "Tools such as ChatGPT threaten transparent science; here are our ground rules for their use," Nature 613 (2023): 612.