The manuscript is written in a way that nobody else could read it suggests that it is preserving knowledge that someone wanted to keep secret.
>> Imagine a book that has beaten the best codebreakers of two world wars, the cryptographers of the Cold War, and generations of linguists.
A book whose pages are covered in elegant flowing writing that no living person can read.
And now imagine that a machine has started to read it.
Over the past decade, headlines have announced that artificial intelligence has found a language hidden in the Voynich manuscript.
They have said it is Hebrew.
They have said a first sentence has finally been translated.
And a darker idea has followed close behind, that this book was encrypted on purpose to hide knowledge so dangerous that someone went to extraordinary lengths to bury it.
If that’s true, there’s a question we can’t avoid.
Should we be opening it at all? I want to take that question seriously, but I also want to do something the headlines rarely do, >> >> which is to follow each claim back to the people who made it, and to the people who disagree.
Some of what you’ve heard is real and astonishing.
Some of it was oversold, and two studies from the last year have changed the picture in ways almost no one has covered.
So stay with me.
By the end, you’ll know what a machine has actually done to this manuscript, what it hasn’t, and why the most unsettling part of the story may not be what’s inside the book.
The book that would not speak.
Let’s start with what we can say for certain.
The Voynich manuscript is a handwritten book of roughly 240 pages made of vellum, small enough to carry in one hand.
It is written in an alphabet that appears in no other document in history, in a language that nobody has ever identified.
For a long time, even its age was a matter of opinion.
Then, in 2009, a team at the University of Arizona radiocarbon dated the vellum.
The result was clear.
The animal skins were prepared between 1404 and 1438.
The book belongs to the early 15th century.
Some summaries repeat an earlier century date, but the laboratory result is the one that counts.
That date does something important, and it comes up again later.
It rules out one of the most famous theories ever proposed.
For years, people suggested that the author was the 13th century English friar Roger Bacon, but Bacon died in 1292, more than a century before this vellum existed.
Where has the book been? Its modern story begins in 1912, when a Polish-born book dealer named Wilfrid Voynich bought it from a Jesuit college at Villa Mondragone, near Rome.
But, there are earlier chapters.
In the 17th century, it was in Prague, in the hands of an alchemist named Georg Baresch, who wrote to the famed Jesuit scholar Athanasius Kircher in 1639, begging for help decoding it.
Later, it passed to a physician, Johannes Marcus Marci, who sent it to Kircher, too.
Kircher, one of the most celebrated polymaths of his day, apparently could not read it, either.
Today, the manuscript sits in the Beinecke Rare Book and Manuscript Library at Yale, cataloged as MS 408.
And what makes it so hard is that it doesn’t behave like a hoax or a code in any obvious way.
The writing is fluent.
The pages are carefully organized.
The text has patterns that feel like language.
And yet, for 600 years, the pages have stayed silent.
But, that silence may be breaking, or so we’re told.
The machine arrives.
There’s a reason people were hopeful that artificial intelligence might succeed where humans failed.
Machines don’t get tired.
They don’t fall in love with their own theories, and they can test millions of possibilities in an afternoon.
Over the last decade, researchers have fed the manuscript’s text into statistical tools, language models, and pattern matching software, and they’ve found things.
One famous study in 2013 by physicist Marcelo Montemurro and Damion Zanette looked at how the manuscript’s words are distributed across the pages.
They found the same kind of statistical rhythm you see in real languages, where certain words cluster around particular topics.
It looked like meaning.
But the story needs a caution right away.
In 2004, the computer scientist Gordon Rugg showed that you could generate text with many of the same features using a simple mechanical method with no meaning at all.
And in a study published in late 2025, the science writer Michael Greshko built a cipher he named Naibi using dice rolls and playing cards that turns ordinary Latin or Italian into glyph patterns resembling Voynichese in frequency, word length, and some grammar-like rules.
He is clear that it’s almost certainly not how the real manuscript was made, but it shows something unsettling.
A manuscript that looks statistically like a language doesn’t have to be one.
So, here’s the problem with asking AI for answers.
A machine is very good at finding patterns.
It is not good at knowing whether a pattern is real.
Give it noise, and it may hand you a beautiful structure that isn’t there.
That’s the heart of the debate, and it’s why the next question matters so much.
Before asking what the book says, we need an easier answer first.
What is this object? A tour of the impossible.
Open the book, and the first thing you notice is how beautiful it is.
The script flows in a steady, practiced hand.
The kind of writing that looks fluent.
Researchers call it Voynich’s.
Whoever wrote it was no beginner, and no one was scribbling nonsense in a hurry.
Then you see the pictures.
The manuscript is conventionally divided into five parts based on its illustrations.
First comes the herbal section filled with drawings of plants.
Some look like real herbs, but many cannot be matched to any species.
And some seem to combine the roots of one plant with the leaves of another.
The second part is astronomical with circular diagrams, stars, and zodiac figures, including recognizable signs such as Pisces and Sagittarius.
The third is the one that stops people in their tracks.
It’s sometimes called the biological or balneological section because it shows crowds of small nude women bathing in green liquid in pools linked by pipes and tubes.
The fourth shows containers and plant parts resembling the pharmacy jars of a medieval apothecary.
And the last is a series of short entries, each marked with a star, which look like recipes or lists.
Here’s a new detail that is worth sitting with.
The vellum was prepared in the early 15th century, and much of the imagery fits that world.
Medieval zodiac conventions, apothecary jars, and bath scenes.
But nothing in the pictures tells us what the text says.
A drawing of a plant could be a real herb, or a drawing of a plant that exists only in the author’s imagination.
And there’s another puzzle.
In 2020, the medieval manuscript scholar Lisa Fagan Davis published a study using digital paleography comparing the handwriting across the book.
Her conclusion was that the manuscript was written by five different scribes.
That’s a big deal.
A lone hoaxer or a lone mystic doesn’t hire four assistants.
This looks like a team project organized and carried out by more than one person.
It’s the layout of a reference book, a guide to something.
Someone invested serious time in cataloging a body of knowledge and then chose to hide it.
One question has trailed the book ever since.
What did they want to keep and who were they keeping it from? To answer that, we need to meet the people who tried to find out first.
A century of dead ends after Voynich’s purchase in 1912.
The book began to draw an extraordinary roster of challengers.
Some of the best known names in the history of code breaking took a swing.
William Friedman, the American cryptologist who helped break Japan’s purple cipher in the Second World War, organized a study group in 1944 to work on the manuscript in his spare time.
Alongside him was his wife and collaborator Elizabeth Friedman, who had cracked rum runner’s codes during prohibition and exposed Nazi spy networks during the war.
The Friedmans found no clear system.
William eventually leaned toward the idea that the text might be an early attempt at a constructed or philosophical language, not a cipher disguising a known one.
That is a very different conclusion from the one often attributed to him, that it was a real language that simply couldn’t be identified.
British code breakers tried, Cold War analysts tried, early computers tried, and every wave of new technology brought a fresh burst of optimism followed by the same wall.
Meanwhile, theories multiplied.
Some people believed it was the work of Roger Bacon until the radiocarbon date ended that idea.
Some suggested it was shorthand or a lost European dialect or a made-up language built to sound real.
Others turned toward the stranger options, a message from beyond Earth, or the visions of a mystic.
One of the most persistent ideas is that the whole thing is a hoax, and here I need to correct something that gets repeated a lot.
Carbon dating did not disprove the hoax theory.
What it did was rule out the idea that Wilfrid Voynich himself forged it, because the vellum is four centuries older than his purchase.
But a medieval or Renaissance hoaxer, someone who intended to sell an impressive, mysterious book to a wealthy collector, remains possible.
The vellum tells us when the animal skins were made, not why someone wrote on them.
There was one more strand of thought, and it’s worth pausing on.
As the failures piled up, some researchers began to wonder whether the problem wasn’t their method, but their assumptions.
What if the text was never supposed to be read in the ordinary way? What if the meaning lived in the pictures, the repetitions, the rhythm? That path led into semiotics and anthropology, and it also ran aground.
The manuscript stayed where it was, patient and unmoved, which is why, when a computer scientist announced in 2018 that he and his student had a language match, the world paid attention.
The Hebrew claim.
In early 2018, a press release from the University of Alberta sounded like a breakthrough.
Computer scientist Grzegorz Kondrak and his student Bradley Hauer had pointed machine learning at the Voynich text.
The paper appeared in the journal Transactions of the Association for Computational Linguistics.
Here’s how it worked.
They took 380 translations of the Universal Declaration of Human Rights, a document that exists in a huge range of languages, and They their algorithm which languages patterns most closely resemble the manuscripts.
The researchers had actually expected Arabic.
Instead, the algorithm pointed to Hebrew.
Then came the second assumption.
They guessed the author had used alphagrams, every word’s letters sorted into alphabetical order, which hides the original word.
They also assumed vowels had been dropped as they often are in Hebrew writing.
Using these steps, they reported that more than 80% of the resulting words could be found in a Hebrew dictionary.
And then came the sentence.
They took the manuscript’s first line, fixed some spellings, ran it through Google Translate, and got “She made recommendations to the priest, man of the house, and me, and people”.
It sounds almost like speech.
And it spread around the world.
But what the headlines missed was how carefully the researchers framed it, and how loudly others pushed back.
Kondrak and Howard themselves said the translations were rough, and that the sentences didn’t always make sense.
They never claimed to have deciphered the book.
Critics were sharper.
The Hebrew-speaking data scientist Argamon pointed out that finding that a text looks more like Hebrew than like other languages is not necessarily saying all that much.
Many scholars also objected that the researchers had used modern Hebrew when a 15th-century European author would have known a very different form.
And Argamon criticized the use of Google Translate, noting that a tool like that will happily produce sentence-like output from nonsense input.
That is the key difficulty.
When you let an algorithm rearrange letters and then check a dictionary, you’ve given it a great deal of freedom.
Hebrew words are short and built from a small set of consonants.
Given enough freedom, almost any string of symbols will match something.
So, where does that leave us? The Hebrew idea is a hypothesis and an interesting one.
It is not a translation.
The logic of hidden words still.
I think it’s worth understanding why the idea was tempting because it reveals something real about how a medieval cipher might work.
An alphagram is a simple trick.
Take a word and sort its letters alphabetically.
The word stone becomes enast.
It doesn’t produce random gibberish.
It produces something that looks orderly but cannot be read at a glance.
For a human being, that’s a dead end.
If you don’t know that the words were alphabetized, there’s no reason to try it.
But a computer can test every possible rearrangement of every word, compare the results against a language, and see what lines up.
That is the strength of what the Alberta team did.
They showed, in principle, how a machine can test hypotheses that humans rarely would.
It’s also the weakness.
If a machine can try thousands of rearrangements for each word, then the chance that some arrangement matches a dictionary is very high, whether the underlying text means anything or not.
This is called a false positive and it’s the central risk in all of this work.
There’s a rule of thumb among code breakers that applies perfectly here.
A real decipherment doesn’t just produce readable words.
It produces readable text that makes sense in context, that can be predicted, and that other researchers can reproduce.
By that standard, nobody has yet passed the test.
The most we can say is that the machine suggested a hypothesis worth checking.
And checking is exactly what scholars have been doing.
Not with algorithms, but with something far more old-fashioned.
They’ve been looking at the pages themselves.
The alphabet in the margin.
In 2014, a team called the Lazarus Project based at the University of Rochester took multi-spectral images of the Voynich manuscript.
This is a technique that photographs a page under many wavelengths of light revealing ink that has faded, text that was erased, and marks the naked eye can’t see.
10 of the manuscript’s pages were scanned.
For years, those images sat largely unexamined.
Then in 2024, the medieval manuscripts expert Lisa Fagin Davis went back to them.
And on the first page, in the right-hand margin, she found something that made her gasp.
She later wrote that she may have shrieked.
There were three columns of letters.
Two were in Roman script.
One was in Voynich characters.
They were lined up as if someone had been building a key matching the strange symbols to the letters of the Latin alphabet.
Who wrote them? Davis compared the handwriting with that of people known to have handled the manuscript.
And she concluded that it best matched the hand of Johannes Marcus Marci, the Prague physician who owned the book in the 1660s.
This matters for a few reasons.
First, it’s the first physical evidence that someone sat down with the manuscript and seriously tried to crack it long before the age of computers.
Marci had famously sent the book to Kircher in 1665.
This suggests he may have worked on it himself first.
Perhaps trying out substitution ciphers or perhaps developing a cipher of his own using Voynich’s letters.
Davis is careful to say that his purpose is unknown.
Second, the faded text in the scans suggests that there is more to discover.
Davis has said as much.
And third, it shows how a century of the best technology still has blind spots.
Those scans had been available for 10 years.
The most important discovery from them came from a human being with deep knowledge of handwriting looking carefully at a margin.
A few popular retellings have also described this as a team breakthrough that confirmed the AI results.
It did not.
The marginalia say nothing about Hebrew and they don’t translate a single word.
They show that a 17th century physician believed the text could be read and that brings us to the next problem.
If the evidence is pointing in two directions, which language is it? Hebrew, Latin, or neither? The Alberta algorithm leaned toward Hebrew.
The margins, by contrast, were written in the Roman alphabet and the book was in the hands of a Latin reading scholar in Prague.
So, some people have asked whether the answer might be both.
It’s not an absurd idea.
In 15th century Europe, Latin was the language of universities, medicine, and law.
Hebrew was studied by theologians, by some physicians, and by people interested in mystical traditions.
A scholar fluent in both wouldn’t have been unusual, particularly in the cosmopolitan trading cities of northern Italy, where many scholars believe the manuscript was made.
A dual-layered code with one language hiding another is possible in principle, but I want to be careful.
There’s a big difference between possible and supported.
So far, the marginalia tell us what a later reader tried.
They don’t tell us what the original author wrote.
Other proposals have surfaced as well.
In 2019, a researcher named Gerard Cheshire published a paper claiming that the text was a form of early Romance language.
The paper was met with fierce criticism from medievalists, including Davis, who rejected the claim.
It’s a useful reminder that in the Voynich field, a confident announcement is almost always followed by experts taking it apart.
What do the specialists believe today? Most of them fall into one of three camps.
Some think the text is a cipher of a real language.
Some think it is an invented language, perhaps a philosophical one of the kind Friedman suspected.
And some think it’s meaningless, an elaborate piece of medieval artifice.
The Dutch researcher Rene Zandbergen, who runs the best known Voynich research site, has described himself as essentially undecided on whether it carries meaning or is a hoax.
That honesty is rare in popular coverage, and it matters because without knowing which of those three worlds we live in, we can’t know whether a translation is even possible.
And that makes the next question more intriguing and more uncomfortable.
If the book does contain real information, why would anyone have hidden it? Why hide it? Here’s the idea that gives this story its darkest edge.
Look again at the bathing women, the tubes, the cycles of stars, and the lists of plants.
A popular theory says that this is a women’s health manual, a guide to fertility, menstruation, or even early methods of birth control.
And because medieval society was uneasy about that kind of knowledge, the argument goes, the author encrypted it for safety.
It’s a gripping story.
It has a hidden book, a powerful church, and a risk of persecution.
And I want to be fair to it because it’s a legitimate theory.
Historians of medicine have documented that women’s healing knowledge in medieval Europe often circulated outside official channels, passed between midwives and healers.
But I also need to be honest about how much we don’t know.
First, nobody has shown that the bathing section concerns reproduction.
The pictures could depict baths for therapy, a common medieval practice, or alchemical processes, or something else entirely.
The imagery alone can’t settle it.
Second, the notion that the church would have persecuted someone for owning a medical book is not well supported.
Medieval medical writing was widely available, and many physicians were trained in church-run schools.
Books of herbal medicine, astrological medicine, and even natural magic circulated in large numbers.
What did draw trouble in some periods was heresy or sorcery.
And the evidence that this book dealt with either is thin.
Third, there are plenty of other reasons to encrypt a text.
A physician might protect trade secrets.
An alchemist might guard a recipe.
A scholar might simply enjoy a puzzle.
In the early modern period, secrecy was a status symbol, a mark of insider knowledge, and a way to make a manuscript valuable.
And then, as always, there’s the possibility that there was no secret at all.
If the book is a hoax, or an invented language, then who was it hiding from has no answer.
So, when the story says someone buried this knowledge for a reason, it’s telling you what it wants to be true.
The evidence says we don’t know, but the idea has an emotional power, and it points toward a bigger fear.
That fear drives the most dramatic version of this story, the danger of decoding.
Here’s the scenario that gets people uneasy.
Suppose a machine does decode the manuscript.
Suppose the contents describe forbidden remedies, rituals, or practices that someone worked very hard to conceal.
Would releasing them be dangerous? It’s a good question, and it deserves a real answer.
Let’s start with the facts.
So far, no AI system has produced a translation that scholars accept.
The single sentence everybody quotes was produced by running rough output through a general-purpose translation tool.
Beyond that, the recurring patterns and ritual language that circulate online come from casual summaries, not from published reviewed analysis.
That means the scariest claims have almost nothing under them.
We are not on the edge of releasing a forbidden text.
We are on the edge of a very human error.
Here’s the genuine danger, and it’s not what you’d expect.
It’s the false positive.
Imagine an algorithm gives you a confident-sounding paragraph, and people repeat it.
A line about priests turns into a story of secret rituals.
A word for plant turns into a claim about forgotten medicine.
Soon, an invented history is circulating as fact, and no one can trace where it started.
We have seen this before.
The Voynich field is full of confident decipherments that collapsed within weeks, and there’s a modern twist.
Today, a plausible-sounding translation can reach millions of people in hours.
There’s also a quieter, more practical risk.
Medieval remedies were often useless and sometimes harmful.
Plants that appear in old herbals can be toxic.
If someone were to treat a half-understood translation as a recipe, the consequences could be real.
That’s why the most thoughtful researchers talk less about forbidden knowledge and more about method.
They ask for transparency, replication, and peer review, and for cultural historians to be involved from the start, not brought in at the end.
And the field has just received two reminders of how strange and unfinished this work still is.
Two new clues.
The first came at the very end of 2025.
In the journal Cryptologia, Michael Greshko published the Naibi cipher I mentioned earlier.
It’s named after a 14th century Italian card game, and it builds Voynich-like text from dice rolls and card draws.
Using it, he could turn ordinary Latin or Italian into glyphs with frequencies, word lengths, and some grammar-like rules that closely resemble the real manuscript.
He’s explicit that his cipher almost certainly isn’t how the manuscript was made.
It approximates Voynichese rather than copying it, and it doesn’t match the way words are placed in lines and paragraphs.
But, he hopes it will become a benchmark, a way to test which features of the manuscript are easy to produce and which are not.
The differences may reveal how the real text was actually created.
The second clue is even fresher.
Just days ago, in October 2026, Lisa Fagan Davis and Colin Layfield, an information systems researcher at the University of Malta, published a paper in the journal Digital Medievalist, arguing that the manuscript was bound in the wrong order.
Their evidence comes from several directions.
The handwriting of different scribes alternates in ways that make no sense in the current page order.
Some illustrations start on one page and continue on another, even though many pages separate them.
And Layfield used a text analysis method known as latent semantic analysis, which compares blocks of text, and found that some neighboring pages are strikingly dissimilar, while distant pages are strikingly alike.
Their proposal is that the manuscript may originally have been a stack of loose double sheets, not a bound book, and that it was assembled later in an order the author never intended.
Think about what that means.
For over a century, every statistical analysis assumed the pages were in their correct sequence.
If they aren’t, a lot of earlier work, including patterns that people took for meaning, may have to be reconsidered.
Davis was careful here, too.
She said the work doesn’t get us to a translation.
She said, in effect, “We don’t know if it’s a code”.
But she added that the finding moves the ball.
That is what real progress looks like in this field, not a sudden decoded paragraph, but a patient narrowing of the possibilities, which brings us to the question the whole story has been building toward.
Should we decode it? So, let’s step back.
Here is what the evidence supports.
The manuscript is a genuine 15th century object made of vellum from between 1404 and 1438.
It was written by several hands in a fluent script with a deliberate structure.
A 17th century physician tried to crack it and left the evidence in the margin of its very first page.
And modern machine learning can test hypotheses that no human could, including that the text is an anagrammed form of Hebrew.
Here is what it doesn’t support.
That AI has decoded the manuscript.
That a first sentence has been reliably translated.
That the book contains dangerous or suppressed knowledge.
That carbon dating killed the hoax theory.
Or that the church buried it.
Each of those claims is either unproven or contradicted.
And here’s the part I find most striking.
The scariest thing about the Voynich manuscript isn’t a secret it might contain.
It’s how easily we can persuade ourselves that we’ve found one.
The human mind hunts for meaning.
Machines hunt for patterns.
When the two meet in front of a mysterious book, you get a perfect storm.
A pattern that looks like a sentence.
A sentence that looks like a story.
And a story that looks like a warning.
It’s the same process that turns clouds into faces.
Only this time, it’s wearing a lab coat.
So, should we decode it? Of course we should.
The book is a window into the 15th century, into the minds of people who wrote, copied, and hid things.
Whether it turns out to be a medical text, an invented language, a puzzle, or a prank, each answer would teach us something real about the people who made it.
But, we should do it the slow way, with open data, with skeptics in the room, and with the humility to say “We don’t know yet”.
A translation we can’t trust is worse than no translation at all.
And whatever the book finally says, I suspect it will be more human than we fear.
Not a warning, just the voice of someone six centuries ago who sat down at a desk, picked up a pen, and wrote something they wanted to keep.
Now, I want to hear from you.
If you could read one page of this manuscript, which would you choose? The plants, the stars, or the bathing women? And do you think it holds a real secret, or the world’s longest-running puzzle with no solution? Drop your theory in the comments.
If you enjoy stories where we follow the evidence wherever it goes, like this video, subscribe, and click the next one on your screen.
Because the strangest mysteries are usually the ones that turn out to be about us.
