If your first thought on seeing the letters "AI" is "then I can ignore this", you are in good company, and for this job you would be right to be wary. A single AI, asked to read a difficult page, hands you a fluent, confident answer that may be wrong, and nothing on the screen tells you so.
That is why I did not ask an AI for the answer. I put five machines (four models, built by three rival companies), in front of a 532-year-old printed book (so that's the 15th Century): the 25 pages of the bookkeeping treatise, and made them read it separately before their outputs were compared. A trial on ten pages from elsewhere in the 630-page book has been run. For the purposes of this essay they are one thing: a technology debating with itself. Claude and I looked at where the machines agreed, where they split, and where they were confidently and instructively wrong, and then further AI instances, looking at cut-out pages of the print, acted as referees, with a few calls from me as the accountant.
This is a novel use of this new technology (as far as I know), and what has come out of it, mistakes included, is as much the story as the book; which is my intention, literally to see what it can do. Read it as a report from an experiment, not some sort of advertisement.
In 1494 a Franciscan friar called Luca Pacioli published a very large book in Venice on arithmetic, geometry and commercial mathematics. Tucked inside it, in thirty-six short chapters, is the first ever printed description of double-entry bookkeeping: how a merchant should keep his inventory, his journal and his ledger, and why every entry must appear twice, when you understand then really it's just common sense. Accountants have been quoting that section for five centuries. Few of them have read the original - up to now, me included.
Pretty much all (non academic) have read only translations, and translations are of course usually interpretations. My own first draft made the point for me. The machines rendered ragioniere as "accountant" until lexical and contextual evidence pointed to something nearer a reckoner; they rendered ditta as "draft", though its meaning remains contested and I now read it as "on credit" (ed: it depends which context!) ; and the 1878 Italian edition used for our earlier translation checks changed a line about what a merchant needs into one about the market. Nobody was careless. The text is hard, and every editor, myself included, has to decide.
So the machines are trying something different. Before anyone translates a word, I want a faithful, checkable reading of the print itself, built in layers so that what the printer set is never confused with what a translator supplied. In modern life I want to see the documents, not listen to someone's description of them, because they are never the same. This essay is about the first layer, and about what was discovered when I asked several Gen AI models to read a book that is 532 years old - indeed one of the challenges is they were all trained on these translations so they have to be told not to use that memory.
Printing reached Europe around 1450, and a book of 1494 belongs to its first half-century. Books from those early years imitate handwritten ones. Pacioli's printer used Gothic type, with long tall s letters that look like f, a curved r that looks like a numeral 2, and capital letters so ornate that C and E are nearly twins. Above all, the pages are full of abbreviation marks.
They look like accents, but they are not. A small wavy line over a vowel means "an n or m has been left out here", so cõmo is commo. A p with a stroke through its tail means per; with a curl, pro; with a tilde, pre. A figure like a 7 is the old shorthand for et, and a 7 followed by a little c is simply etc. Medieval scribes saved parchment and time this way, and early printers copied them, partly for the look and partly because abbreviations let a compositor stretch or squeeze a line to fit. A large share of the words in the treatise carry one of these signs.
Computer text recognition is trained mostly on later printing, where such signs are rare. That is why a machine can read a book printed in Venice in 1494 and write nonsense.
We are working from a high-resolution (300 dpi) scan of the Library of Congress copy. The bookkeeping treatise runs to 25 pages of it. We ran five independent machine passes using four models on those pages, each from the images alone and each blind to the others, following the same written rules for transcribing exactly what is printed, abbreviations included, without tidying: Claude Sonnet 5.5 from Anthropic; GPT-6-Sol at two levels of effort and GPT-6-Astra at its highest, from OpenAI; and Gemini 3.1 Pro from Google. We also ran a specialist open-source program for early printed books, Kraken, as a reference rather than a voter.
Then we compared them, word by word.
The first surprise is how much. Of 28,775 word and punctuation tokens in the treatise, 76 per cent were identical across all five machine outputs. Agreement alone does not establish accuracy. On letters they agree much more closely, that means the abbreviations less so.
The second surprise is where they differ. The disagreements are not scattered; they cluster in the hard places: worn type, ornate capitals, and above all the abbreviation signs. In 4,628 places at least one machine output differed. In 1,417 of them there was no majority at all: no single version was written by as many as three machines. These are the words where a single AI would have given a smooth, confident, and possibly wrong answer, and where nobody would ever have known. This is the very purpose of this project, to show that like us humans, machines do better when they too work as a team. And as an aside I have found when the AI starts talking rubbish it is because us humans have designed in memory loss, much like Alzheimers (in my sad experience) . It's to do with the cost of computer memory, but it can destroy carefully won progress as happened several times in this very project, in my opinion without close supervision this renders Gen AI incapable of reliable use - Anthony
So this is the point of using several. Each machine has its own idiosyncrasies. One marks far fewer doubtful characters per page than another. One writes a certain sign as ꝗ while the other four write q̃, and it is almost certainly the same mark. It is the Big-Endians and Little-Endians of Gulliver's Travels again: a quarrel over which end of the egg to crack, between AI models that are each perfectly sure of their end.
The machines also found their own mistakes, which is the very idea. On one page, one of the AI models, Claude, read the etc. sign as a round r followed by a c, many times. Another AI model had it right. A side-by-side picture of that sign next to a genuine round r from the same page settled it in a moment: one is shaped like a 7, the other like a 2. When the same AI model later read single words cut out of the page, without a sentence to lean on, it got the sign right every time in the sample. Habit and memory had been misleading it in context.
The abbreviation marks turned out to be the weakest layer by far. The AI models can agree on a letter and still disagree on which little mark sits above it. They often differ over abbreviation marks, including the per sign. That suggested a better order of work: settle the signs first. A substantial share of the recorded differences involved marks or glyph conventions, but that count does not tell us how many a signs-first method would resolve correctly. I regret not spotting this first as much of the earlier work would have been much more accurate instead of having to "go round again" to do all the work again.
Because the marks were the weakest layer, a good deal of the work went into them. Across the whole treatise, about one word in seven carries a sign: 419 different abbreviated forms among the words all five machines agreed on. We made a key to the little signs, with photographs cut from the book showing each mark and what it means. Then we tested our reading of every sign against two later printings that spell most words out, Gitti’s 1878 edition and the 1523 reprint. We worked backwards from their completed words to the signs in the original, and asked whether our reading of each sign gave the same word. In modern parlance we "reverse engineered" the abbreviations. Where it did not, we went back to the page image. That check found about thirty corrections, most of them a missing or wrong sign, among them nineteen missing strokes on partita. Others came from different checks: a blind second opinion by two of the machines, and, most recently, a blind recheck of the doubtful places, which restored four more strokes.
A later printing is strong evidence of how a sign was understood, but it is not proof. The 1523 compositor worked from a copy of the 1494 book and could misread it too, and Gitti modernised as he went. We also ran a trial on ten pages from elsewhere in the 630-page book in which the machines listed the signs first and read the words afterwards. That trial has not been extended to the whole book, and its results are not reported here.
Our reading of the signs agreed closely with how Gitti and the 1523 compositor expanded them; the exceptions were mostly a few signs the machines all missed. Everyone can be wrong, machines and editors alike, and that is where human challenge comes in: someone who can read a mark better than we did should say so. Personally I prefer the 1523 version because it is close to the time of the original and I think the more accurate one.
I wish we had done this first. Had the signs been settled before anyone read a word, much of the disagreement between the machines, and much of the later checking, would have been aimed at the right place from the start.
This essay reports the first reading stage, the machine readable version of the 1494 print. The work has then gone further: from that 1494-computer-readable-version we have made modern Italian and English versions, as machine created drafts. The 1523 reprint and Sangster and Santini’s 2022 translation of Chapter 1 supplied later comparisons; the Editorial Method page explains how the 1878 Italian edition was used. Neither the resolved reading nor the translations has been checked line by line by an expert in the period. A handful of relatives are already helping, answering a short list of the disputed words from a private page: a close-up of the print and a few numbered suggestions, and a pick. Their eyes, it turns out, are among the most valuable resources we have.
The strangest lesson is a modest one. If you ask a single AI to read a difficult page, you get an answer in a confident voice and no sign that its neighbour would have said something else. Disagreement is information. Wars have started over less than a difference of convention, and the cure is the same here as in Lilliput: put the eggs side by side. I think Kraken is an unsung hero in all this. It is a neural-network text-recognition program rather than an LLM (large language model), trained on printed books like this one.