For more than a hundred years, the Voinich manuscript has beaten every codereaker who tried it.
Nobody has read even one page of this book from the early 1400s, sitting in a vault at Yale.
Even William Freriedman, the father of American codereing, gave it decades of his life and got nowhere.
Then in 2018, a university computer lab, not a single historian among them, turned an AI loose on the book.
What came back looked like the first readable sentence in 600 years.
And within days, experts around the world were fighting over what it said.
The book that broke every codereaker.
I’m Ambrose Wild.
And if a headline about AI cracking this manuscript is what brought you here, you’re standing exactly where millions of readers stood when this story first broke.
And by the end, you’ll know whether that headline fooled you.
So before that sentence, let’s put the actual book on the table because every fight that came later started with the object itself.
At Yale’s Bine Rare Book and Manuscript Library, it’s cataloged as MS48.
It runs about 240 pages of calfskin vellum, small enough to hold in two hands, and it’s covered in more than 35,000 words of looping script that matches no alphabet anyone has ever identified.
In 2009, a team at the University of Arizona cut four tiny samples from its pages and radioarbon dated them.
The results pointed to one narrow window between 1404 and 1438.
That same year, chemists at Mcronone Associates tested the ink and paint and found nothing a medieval workshop couldn’t have made.
Hold on to that window of years.
The book isn’t even complete.
Someone numbered the pages long after they were written and those numbers skip in places which tells us several leaves have been lost along with whatever was on them.
Then there are the drawings.
There are plants that no botonist has matched to a real species and a wheel of zodiac signs with two of them missing, probably lost along with a tornout page.
Naked women sit in green pools linked by strange pipes.
A giant foldout opens into nine connected circles, and shelves of jars look like they came from an apothecary.
Around the zodiac wheels, somebody later wrote month names in ordinary Latin letters in what looks like a southern French dialect.
Those few words added by a different hand are the only lines in the whole book anyone can actually read.
The script itself has habits that feel wrong.
The same word will often appear two or three times in a row.
Like cuc ordinary writing in any known language almost never does.
Scholars broadly agree on everything I’ve just described and the agreement ends the moment anyone tries to read the words.
Freriedman found that out the hard way.
In the last years of World War II, he gathered a volunteer team of wartime codereers to study the book.
And in the 1960s, he tried again with another group.
Neither team cracked it.
So Freriedman did something strange with his best guess.
In 1959, he hid it inside an anagram printed as a footnote in a scholarly article about Shorser and sealed the solution in an envelope.
That envelope was opened after he died.
His answer was that the book was an early attempt to build an invented universal language from scratch.
Keep that picture of a codereaker hiding his answer in scrambled letters because the AI that came next assumed the entire manuscript had been written the same way.
The lab that went hunting for a language.
Greg Condra builds language software for a living.
He’s a computing science professor at the University of Alberta and in 2016 he and his graduate student Bradley Hower published a paper that set out to answer a smaller question first.
Which language might be hiding underneath the script.
Their training material was the Universal Declaration of Human Rights because the United Nations has it translated into hundreds of languages.
The pair fed their software that document in 380 of them so it could learn the fingerprint each language leaves behind like which letters tend to sit together and how words usually begin and end.
Then they showed it the Voinich text.
They went in expecting Arabic.
Under their strongest method, Hebrew came out ahead of all 379 other languages.
To get there, they made two big assumptions.
The first was that the script drops most of its vowels, the way written Hebrew and Arabic often do.
The second was that every word had been scrambled into something called an alphagram, where the letters get sorted into alphabetical order.
So the word cat would be written as a C t.
If that sounds familiar, it should because it belongs to the same family of tricks Freriedman used to hide his own answer.
Now, the software’s job was unscrambling.
It took each Voinich word, treated its letters as a sorted pile, added vowels back in, and searched for a Hebrew word that fit the sentence that made the news.
The team ran that process over the first 10 pages.
More than 80% of the unscrambled words turned up in a Hebrew dictionary.
And on paper, that sounded huge.
But a list of dictionary words isn’t a sentence.
So they asked a Hebrew-speaking colleague, the computer scientist Moshe Cppel, to try reading the opening line.
He couldn’t make it work.
So the researchers corrected a couple of spellings and ran the line through Google Translate.
This is what came out.
She made recommendations to the priest, man of the house, and me and people.
Contra said the sentence was grammatical and could be interpreted.
In January 2018, the University of Alberta published a story about the work and within days it had spread across newspapers, tech sites, and TV news with plenty of headlines announcing that AI had finally decoded the Voinich manuscript.
Read that sentence one more time, though.
After six centuries of silence, the first message out of the world’s most mysterious book was a line about a priest and a household.
It was ordinary and oddly flat, and that flatness sent specialists back to look at how the machine had been allowed to build it.
Three letters that can spell almost anything.
You can test this yourself.
Take three consonants, R, S, and T.
Now follow the same rules the algorithm used.
Put the letters in any order you like and add any vowels you want.
You get star, rats, arts, sar, rust, stir, rest, store, and roast.
That’s nine English words from one scrambled cluster.
And the dictionary gets bigger with every letter you add.
That’s what critics spotted.
When the rules let you reorder letters freely and fill in vowels however you want, most short clusters will match something in almost any language.
So an 80% hit rate on its own doesn’t prove much at all.
Hebrew specialists pushed back too.
The language columnist who writes as Filologos for Mosaic magazine went through the method and argued the text simply wasn’t Hebrew.
Then there was the translator.
Google translate learned its Hebrew from the language spoken in Israel today.
And it was being asked to read words that if the theory were right came from a manuscript written centuries earlier.
And every spelling fix made by hand was a human choice.
The kind of choice that can nudge a result towards something readable.
Condrach himself said a historian of ancient Hebrew would need to confirm any meaning.
None of this was new either.
In 1921, a University of Pennsylvania professor named William Roma Newold announced that he had cracked the book.
He claimed tiny strokes inside each letter formed a microscopic Greek shorthand and that the author, the 13th century English frier Roger Bacon, had described galaxies and living cells centuries before anyone could have seen them.
In 1931, the scholar John Manley examined the same marks and showed they were just cracks in drying ink.
Newold had died in 1926, still certain he was right.
The software had repeated a century old mistake at computer speed.
And that realization is what unsettled the people who study this book.
Even so, the pattern wasn’t finished because the next claim arrived in 2019 with a university’s official seal on it.
The press release a university took down.
In May 2019, the University of Bristol announced that one of its research associates, Gerard Cheshire, had solved the manuscript in about 2 weeks.
His paper published in the journal Romance Studies said the text was written in a lost language he called protoromance, an early ancestor of Italian, Spanish, and Portuguese.
He even named a reader.
He argued the book had been put together by Dominican nuns as a reference for Maria of Castile, Queen of Aragon, and that it covered herbal cures, therapeutic bathing, and women’s health.
Specialists immediately asked an awkward question.
If a whole royal court could read this language, why does it survive in exactly one book written in an alphabet found nowhere else?
The story went around the world within hours and within about a day it was falling apart.
Lisa Fagan Davis, the executive director of the Medieval Academy of America, pointed out that the method let the same symbol stand for different letters whenever that produced a better word.
And she described the reasoning as circular.
Bristol took its news story down and said concerns had been raised about the research.
And Bristol wasn’t even the first famous myth of that stretch.
In September 2017, the Times literary supplement ran a feature by the researcher Nicholas Gibbs, who said the book was a Latin medical guide written in abbreviations.
Scholars pointed out almost immediately that the Latin he produced wasn’t grammatical.
So, in under two years, three separate solutions hit the headlines.
One built on abbreviations, one on AI, and one on a lost language.
And each fell apart soon after it appeared.
Line them up, and you notice what they share.
Every method gave its inventor so much freedom that it was bound to produce words.
That raises a darker question.
What if nobody can read the book because there’s nothing written in it at all?
The case for a book with nothing inside.
In 2004, Gordon Rug, a researcher at Keel University, argued exactly that in the journal Cryptologia.
He showed you could produce voinich style text with a card and grill, a card with small windows cut into it.
You slide the card over a big table of word beginnings, middles, and endings.
Copy whatever shows through the holes, and you get page after page of words that look like a language and mean nothing.
Rug built his own tables and produced voinage style text this way to show the idea actually works.
It even explains the book’s strangely even word lengths.
Normal languages lean heavily on tiny words you barely notice, but Voinich words cluster around the same middle length.
The grill is named after Geralamo Cardano, the Italian scholar who described it in 1550.
Now think back to the vellum.
Those [clears throat] four samples from Arizona put the parchment between 1404 and 1438, more than a century before the tool was written about.
Rug’s reply is that simpler grills could have existed earlier, or that a hoaxer used old blank vellum much later.
It’s a genuine gap and nobody has closed it.
Then the gibberish camp found a way around it.
In 2020, Torston Tim and Andreas Sha published a simpler idea in the same journal.
A scribe doesn’t need a grill.
They copy a word from a line above, change a glyph or two, and write it down again and again.
That method needs nothing but a pen and patience, and it explains the habit that felt so wrong on the first pages.
Those words repeating two and three times in a row.
It’s a strong case, but the numbers inside the text point somewhere else entirely.
The text that behaves too well to be nonsense.
If you’re enjoying this kind of deep research, subscribing is the easiest way to catch the next one.
Before any of this can be measured, someone has to turn the handwriting into letters a computer can count.
In the 1990s, volunteer researchers built a standard system for that called the European Voinich alphabet.
Even that step involves judgment calls like whether a certain loop is one letter or two.
And every choice quietly shapes the statistics that come out the other end.
With that caveat, count every word in the manuscript and rank them from most common to least and you get the curve found in almost every human language where the top word appears about twice as often as the second and three times as often as the third.
It’s called Zip’s law and random scribbling doesn’t usually produce it so neatly.
The letters are also highly predictable.
Researchers measure that with something called entropy.
And the Voinich script scores lower than almost any European language, meaning each character strongly hints at the one that follows.
Some glyphs only ever appear at the start of a word and others only at the end, like spelling rules.
In 2013, the researchers Marcelo Montimurro and Damian Zanette ran a keyword test in the journal plus one.
In real books, topic words bunch up in the chapters where they belong.
The Voinich text does the same thing.
Certain words cluster on the plant pages and others on the bathing pages as if the writing is describing the pictures beside it.
Back in the 1970s, the US Navy cryptologist Prescott Courier noticed the book seems to contain two dialects which he labeled A and B.
Each has its own spelling habits and each lines up with different handwriting.
Courier also found that the first and last words of a line behave differently from the rest, as if each line were a unit of its own, the way entries in a list or a table would be.
And in 2025, the science journalist Michael Gresko published a cipher in cryptologia that uses dice and playing cards, things a 15th century Italian could easily have owned to turn plain Latin or Italian into text that looks remarkably like voinich writing.
He called it the naive cipher.
After an old Italian card game.
It shows that a genuine message could have been hidden this way within the right window of years, using nothing a scribe couldn’t buy.
So, a lost language, a clever cipher, and pure nonsense each fit part of the evidence.
And all three camps quote the same statistics.
To break that tie, you’d want to know where this book has been and whose hands have held it.
That record has holes of its own, the owners, and the gaps between them.
The earliest name we can actually see is on the first page.
Under ultraviolet light, a faded signature appears belonging to Jacobus Horsiki Deepenek, the chemist and herbal doctor to the Holy Roman Emperor Rudolph II in Prague.
The title in that signature is a clue in itself because he was only granted it in6008.
So the book was in his hands sometime after that.
That same page hides something else.
In 2024, when Davis studied multisspectral images of it, she found a faint column of Latin letters running down the margin.
It looks like an early reader trying to build a key of their own and never finishing.
Rudolph is supposed to have bought the manuscript for 600 gold duckets, believing it was Bacon’s work.
But that story comes secondhand from a letter written decades later, and no payment has ever turned up in the imperial records.
By 1639, the book belonged to a Prague alchemist named Gaog Barish, who wrote to the Jesuit scholar Athanasius Kercher in Rome, pleading for help reading it.
He picked Kercher for a reason.
Kcher was famous across Europe for claiming he could translate Egyptian hieroglyphs.
We now know his translations were almost entirely wrong.
So, the first expert this book ever reached already had a record of confident decodings that later turned out to be wrong.
After Barish died, his friend Johannes Marcus Marci, recctor of the university in Prague, sent the book to Kercha in 1665 with a covering letter.
That letter was still tucked inside when the manuscript resurfaced.
Then the trail goes dark for about 250 years.
In 1912, the rare book dealer Wilfred Voyage bought it from the Jesuits at the Villa Mondreon near Rome, who were quietly selling off manuscripts to raise money.
For years, Voinich wouldn’t say where he’d found it.
And of everyone in that long chain, he was the one with the most to gain from making the mystery bigger.
The dealer with a motive.
His own past sounds invented.
As a young Polish revolutionary, he was arrested by the Russian authorities, sent to Siberia, and escaped to London in 1890, where he rebuilt his life selling rare books.
He also promoted the Bacon story hard and New Bold’s 1921 decoding gave that story scientific cover.
A lost work by Bacon would sell for a fortune and Voyage was the man selling it.
So some writers have suggested Voyage forged the entire manuscript himself on old blank vellum sometime before 1912.
This is where the evidence turns against him.
The radiocarbon window of 1404 to 1438 comes back and this time it has help.
The ink tests found materials consistent with a medieval workshop.
The letters from Bes in 1639 and Marci in 1665 describe a strange illustrated book in an unknown script more than 200 years before Voyage was born.
Then in 2020, Davis compared the handwriting page by page and identified five different scribes.
That’s very hard to square with one man forging it alone in London.
Even the money argues against a scheme.
After Voyage died in 1930, the book passed to his widow and then to the New York dealer Hans Krauss, who bought it in 1961 for $24,500.
Krauss hoped to sell it for far more, never found a buyer, and in 1969, he donated it to Yale, where it has stayed ever since.
So, the manuscript is genuinely old, genuinely strange, and still unread.
That brings us back to the machine and to why AI keeps failing on this book in particular, what the silence is telling us.
Machines have helped crack lost writing before.
In 2011, a team including the computer scientist Kevin Knight decoded the Copi cipher, a 105page book of secret symbols that turned out to hold the rituals of an 18th century German secret society.
In 2023 and 2024, machine learning read words inside the charred scrolls of Herculanium buried by the eruption of Mount Vuvius in 79 AD without anyone unrolling them.
And those readings were only accepted after paparologists checked the letters one by one.
And back in 1952, the architect Michael Ventress cracked linear B, a Bronze Age script found on Cree and in mainland Greece with pencil, paper, and patience.
Ventress had an anchor, though.
Names of Cret and towns like Nos turned up on the clay tablets, giving him known words to test his guesses against.
In every one of those cases, the language underneath was already known.
German in one and Greek in the other two.
The Voinich manuscript hands you three unknowns at once.
Nobody knows the script.
Nobody knows the language.
And nobody can prove there’s a message.
The only anchor the book offers is that handful of month names around the zodiac wheels.
And those were written by someone else.
That’s why the 2018 sentence matters more today than it did back then.
Anyone can now paste a page of this book into a chatbot and get a fluent, confident translation in seconds.
And every year, another batch of videos and articles claims some unnamed lab has finally cracked it.
Those systems are built to produce text that sounds right.
They’ll produce an answer for a page that has none, and they’ll sound sure about it.
A priest in a household scared nobody.
What shook the specialist was watching software make a silent book speak and watching the whole world believe it for a few days.
So after more than a hundred years the manuscript is right where this story started in a vault at Yale still unread.
Every decoding so far has collapsed under checking.
Every hoax theory has run into that radioarbon date and the statistics keep arguing both sides.
I want to know where you land.
Is it a lost language, a clever cipher?
Or a beautiful fake.
Tell me in the comments and tell me what evidence would change your mind.
I’m Ambrose Wild and there are plenty more unsolved mysteries like this one still waiting on the shelf.
So, subscribe and I’ll bring you the next one.
