Printing and the fixity of the text: why continuous ingestion follows

Printing and the fixity of the text: why continuous ingestion follows

A manuscript is a population, not a document. Every copyist who touches it can correct a word, gloss a difficult passage, or introduce an error that the next copyist may fix or compound. Two manuscripts of the same work, separated by a century of transcription, can diverge in wording, in order, even in argument. Nobody controls this. The text is alive in the bad sense: it drifts, and nothing holds it still.

Set that same text in type and strike off five hundred copies, and something categorically different happens. Every copy carries identical wording. A reader in Basel and a reader in Kraków can compare the same page number, the same diagram, the same footnote, and know they are looking at the same thing. Historians call this typographical fixity: the moment a text stops being a moving population of variants and becomes a single, dispersed, authoritative state. It is not that print is more accurate than manuscript — early printed books are full of errors — but that whatever state gets struck off is reproduced without further drift. The error, like the correct reading, now travels perfectly.

This bought things that manuscript culture could not offer. Citation by page number became possible, because the page was the same page everywhere. Cumulative correction across editions became possible, because a corrected second edition could supersede a first in a way that a corrected manuscript copy could not reliably supersede its exemplar — the fix propagated by replacement, not by chance. Standardisation of reference works, of scripture, of legal codes, followed. But fixity created a new failure mode that manuscript culture, oddly, did not have in the same form: an error, once printed, is no longer a local mistake. It is now everywhere at once, identically, and correcting it means reaching every copy already dispersed beyond recall. The errata slip — a leaf of corrections bound in after the fact, sometimes years after the fact — is publishing's first formal admission that the artefact is frozen while the world, and the printer's own understanding of the text, keep moving.

Where the idea comes from

Elizabeth Eisenstein set out the strong version of this argument in The Printing Press as an Agent of Change (1979): standardisation and preservation of texts, not literacy on its own, drove the scientific revolution, because scholars in different cities could now compare an identical diagram and build corrections on top of a shared, stable base rather than each working from a locally corrupted copy. Adrian Johns answered her in The Nature of the Book (1998), arguing that fixity was never a property that the press simply conferred. It was a social achievement, fought for by guilds and printers against rampant piracy, secured by licensing, by entry in the Stationers' Register, by reputational policing of which shops could be trusted not to mangle a text. Early modern books were pirated, reset with errors, reissued with cancel leaves pasted over botched pages, altered mid-print-run. Fixity, on this account, was a claim made about books, enforced unevenly, not a physical fact about the technology.

The dispute has not been resolved, and it does not need to be for the present purpose. Both historians agree on the thing that matters here: the errata slip existed, was common, and stood as a standing confession that reproduction had outrun correction. Whether fixity was intrinsic to the press or manufactured by institutions around it, readers treated a printed page as authoritative, and were periodically burned for it. In 1631, Barker and Lucas's printing of the King James Bible dropped a single word from Exodus: the seventh commandment came out as "Thou shalt commit adultery." Roughly a thousand copies carried this before anyone caught it. The printers lost their licence and paid a heavy fine; most copies were burned; around a dozen survive today, prized precisely because the error survived along with them. The mistake was not in the text as understood. It was introduced at the point of reproduction, and reproduction then did its job perfectly, distributing the error with the same fidelity it would have distributed the correct reading.

The turn

A model trained on a fixed corpus up to a cutoff date is doing something a historian of the book would recognise immediately. Its weights are a struck-off edition: one state of the world's text, fixed at the moment of training, reproduced identically for every reader who queries it, regardless of what has happened since. Retrieval augmentation, system prompts, periodic fine-tunes — these are errata slips. They are corrections pasted in after the fact, patching an artefact whose body cannot be cheaply re-set. They work, to a point, the way the errata slip worked: imperfectly, belatedly, and only for readers who bother to check the last few pages of the book.

A system built instead to observe a bounded scene directly — a room, a workspace, a live camera feed — does not have this problem, but it does not solve it either. It reports what is in front of it, accurately, now. It is a reading taken from an open page, not an edition. It says nothing about the volumes on the shelf behind it, nothing about last month, nothing about anything not currently in view. The printed page and the live sensor share a defect that has nothing to do with accuracy: neither one carries a revision date. The book asserts its contents as though time had stopped at the printing house. The live scene asserts its contents as though nothing existed outside the frame. Both assert without qualification. Neither holds a belief that can be revised, because neither is structured to say when it was formed or what would overturn it.

The remaining move, on this axis, is to stop treating intake as a single ingestion event — training or observation — and instead keep every relevant stream open, indefinitely, with each incoming claim tagged by source, by timestamp, by what would supersede it. That is not a fourth kind of evidence, arriving after language and after perception. It is the closure of the set: a frozen corpus, a live scene, and the union of every stream still running, held as claims that can be dated and dethroned rather than facts asserted for all time. Publishing took roughly four hundred years to travel from the errata slip to something resembling continuously amended text — consolidated statute with an "as at" date, living systematic reviews that replace their own prior editions on a schedule tied to evidence rather than to print runs. The same distance need not take four hundred years again, but the direction of travel is the same one publishing already walked.

What narrows the claim

The most serious objection is Johns's, aimed at the analogy's root: fixity was never a property of the press, only a claim enforced by licensing and guild policing against real, sustained piracy and misprinting. This does not weaken the argument; it sharpens it. A model's apparent authority is likewise not a property of its weights. It is manufactured by evaluation suites, provider reputation, interface design that presents an assertion in the confident register of a finished book. The argument needs only that readers treat the output as stable and are occasionally injured when it is not. Both halves of that were true of print and remain true here.

A second objection cuts closer to the bone: continuous revision destroys the very thing fixity bought. A citable page, a reproducible experiment, a decision that can be audited later, all depend on text that does not move under the reader's feet. A system that revises its beliefs constantly cannot be quoted, cannot be regression-tested, cannot be held to what it said last Tuesday. This is correct, and it is a real cost, not a rhetorical one. The answer available to publishing was not to abandon fixity but to relocate it — from the artefact to the citation. A repository moves; a commit hash does not. A statute is consolidated and amended continuously; the "as at" date fixes what a particular citation meant at a particular moment. Anything built on continuous intake owes the same discipline: an immutable, timestamped, sourced snapshot of any belief actually acted upon. Absent that snapshot, this objection stands unanswered, and whatever is being described is not what this argument defends.

A third objection questions the mechanism rather than the ethics: an errata slip fixes a discrete, locatable misprint, while staleness in a large trained system is distributed through every parameter, a drifted prior rather than a wrong word on page 412. The disanalogy is real. But it changes which historical failure applies, not whether one does. Distributed drift resembles the superseded edition more than the misprint — the 1911 encyclopaedia is not wrong entry by entry, it is simply from 1911 — and publishing answered that with dating, which is cheap, rather than re-setting, which is expensive. Aviation runs exactly this pairing today: charts fixed on a 28-day cycle worldwide, with NOTAMs — live text streams — carried alongside and trusted over the paper when the two disagree. The cheap move, dating the belief, is the one modern systems most often skip.

What this does and does not establish

The concept establishes that intake has an intrinsic ceiling, not a technical one: a frozen corpus and a live scene are both undated assertion, and the only remaining position is dated, sourced, revisable belief. It does not establish that such a system is easy, cheap, or built. It does not establish that revision replaces fixity wholesale — only that fixity moves from the page to the citation. And it does not establish that intelligence itself has a ceiling, only that this one axis, measured from Gutenberg to the present, has exactly three rungs and no room above the third for a different kind of intake to appear.