Inside a vault at Yale’s Beinecke Library sits the
Voynich manuscript, 240 calfskin pages covered in fluent, practiced handwriting, in a script that
Exists nowhere else on Earth.
In 600 years, nobody has read one sentence of it, and you could
Study every page tonight and do no better.
William Friedman, whose team broke Japan’s Purple
Cipher before Pearl Harbor, spent decades on it and came away with nothing.
The NSA’s
Best analysts picked up where he left off, and they failed too.
So we tried what none of them
Had.
We gave an AI every study of the manuscript, from the carbon dating to the handwriting
Analysis, and asked it one colder question.
Never mind what the Voynich manuscript says.
What is it?
The second thing it found contradicted the first.
I’m Ambrose Wild, and the first thing we handed
The AI was Yale’s full set of page scans, because the object in them is stranger than its
Reputation.
The Beinecke Library’s outer walls are panels of marble cut so thin that daylight
Glows through them, and the Voynich manuscript is one of the few books inside that people
Ask for by name.
It’s roughly the size of a hardcover novel.
The page numbers run to 116,
But only 102 leaves survive, so at some point somebody removed 14 of them, and nobody knows
Why.
A wormhole runs through the first page, which tells you the book once sat against a wooden
Cover that no longer exists.
Open it and you get plants that botanists still can’t agree on, zodiac
Wheels ringed with naked women holding stars, and bathers standing in green pools linked
By tubes that look like plumbing.
One foldout opens into nine connected circles, like a
Map of a country that was never on any map.
And the detail that stopped the AI first
Is this one.
The only words in the book that anyone can read are month names, scribbled
Onto the zodiac pages in ordinary Latin letters by a later owner.
Scholars still argue about
Whether they’re French, Occitan, or something else.
So the only readable words in the Voynich
Manuscript came from one of its later readers.
For most of the 20th century, the easy explanation
Was that the whole thing was a fake.
Wilfrid Voynich, the book dealer who bought it in 1912
From a Jesuit collection at the Villa Mondragone outside Rome, had every reason to hype it, and he
Pitched it as the secret work of the 13th century friar Roger Bacon.
In 1921, William Romaine
Newbold, a philosophy professor at the University of Pennsylvania, announced he’d cracked it.
He
Claimed every letter hid microscopic shorthand strokes, and that Bacon must have invented the
Microscope to write them.
He even read one of the circular drawings as a spiral galaxy, proof,
He said, that Bacon had built a telescope as well.
Ten years later, John Matthews Manly, a Chaucer
Scholar who had broken German spy codes for American military intelligence in the First
World War, looked at the same letters under magnification and showed the strokes were cracks
In dried ink.
Manly also showed that Newbold’s method could pull different messages out of
The same passage, which meant the messages were coming from Newbold himself.
Newbold’s
Reading was finished.
The suspicion that the book itself was a fake stuck around.
Maybe Voynich
Forged it himself.
Maybe it was Edward Kelley, the alchemist and con man who worked the court
Of Emperor Rudolf the Second in the 1580s.
Then, in 2009, a team at the University of Arizona
Carbon dated four samples of the calfskin, and every one landed between 1404 and 1438.
Tests
On the ink and paint found materials that fit a medieval workshop.
That closed the door on a late
Forgery.
Whoever made the Voynich manuscript was working decades before the Gutenberg Bible,
And paying for expensive vellum to do it.
That history, plus every major study published
Since, is the pile we gave the AI.
The first rule was to translate nothing at all.
For every
Study, work out what it should show if the book is a real language, what it should show if it’s a
Cipher hiding a language, and what it should show if the text means nothing at all.
Then score
Every result against all three.
Every finding you’re about to hear comes from published
Research, and the AI’s job was to weigh it, line it up, and look for what connects.
It’s
Hard for any one researcher to keep a century of Voynich papers in view at once, and a machine
Doesn’t get tired or fall in love with its own theory.
What came out was a scoreboard,
And the scoreboard refused to settle.
It started with the most famous test in the file,
Zipf’s law.
In any long piece of real writing, a handful of words do almost all the work.
In English, short words like the and of show up more than any others, and the counts fall
Down a smooth, predictable slope.
Linguists find that same slope in English, Latin, Chinese,
And languages that died a thousand years ago.
Run the Voynich manuscript through the same count
And it sits right on the curve.
For decades, people treated that as proof of a real language.
The AI marked it as weaker than it sounds, because some kinds of nonsense can land on that curve
Too.
Keep that in mind, because it comes back.
The stronger evidence came from where the words
Live in the book.
In 2013, the physicists Marcelo Montemurro and Damian Zanette measured which
Voynich words carry the most information, and found those words cluster in particular
Sections.
Their method doesn’t need to know what a single word means, which is why it’s useful
Here.
The key words gather the way words like leaf and root would pile up in a gardening chapter
And vanish in a chapter about the stars.
In 2021, a team at Yale ran topic modeling, the same
Technique used to sort news articles by subject, and the topics it found largely lined
Up with the pictures.
Herbal pages share vocabulary with herbal pages.
Bathing
Pages share vocabulary with bathing pages.
There’s even a split inside the script
Itself.
In 1976, Prescott Currier, a veteran codebreaker for the Navy and the
NSA, noticed the manuscript is written in two statistical dialects, which he called A and
B.
Certain word endings are everywhere in one dialect and nearly absent from the other, and
The dialects tend to switch where the subject matter does.
Currier also saw at least
Two handwritings, one larger and rounder, one tighter and more cramped, and each
Seemed to travel with its own dialect.
So the first finding looked solid.
Measured across
Whole pages and sections, the Voynich manuscript behaves like a book with chapters about different
Things.
That held right up until the AI zoomed in on a single word, which is where the second
Finding arrived and contradicted the first.
Real languages are loose at the level of letters.
In English, the same four letters give you stop, pots, tops, and spot, and nobody thinks twice
About it.
Voynich letters don’t move like that.
In 2022, at a Voynich conference hosted by the
University of Malta, the researcher Massimiliano Zattera showed that nearly every word in the
Manuscript can be built by filling 12 numbered boxes from left to right.
Each letter belongs in
One box, or at most two.
Boxes can be skipped, and the order almost never runs backwards.
About
87 percent of all the words in the book fit that pattern exactly, and another 10 percent turn
Out to be two of those words stuck together.
Currier had spotted something just as odd decades
Earlier.
The line of text itself seems to matter.
Some letters prefer the first word of a
Line, and others crowd onto the last word, almost as if running out of space on the
Page changed the spelling.
Imagine English words changing shape depending on where the
Margin falls.
The text also repeats itself in ways real prose avoids.
The same word
Often appears two or three times in a row, and nearly identical words sit side by side, like
Writing stop, stops, stop across a single line.
Then there’s predictability.
Linguists can
Measure how easy it is to guess the next letter once you’ve seen the current one.
In the
1970s, the only real language that came close to the Voynich on that measure was Hawaiian,
Which gets by with just 13 letters.
In 2020, Luke Lindemann and Claire Bowern at Yale compared
The manuscript against more than 300 texts in dozens of writing systems.
Voynichese came out
More predictable than every single one.
They tried the usual explanations, like heavy
Abbreviation or leaving out the vowels, and none of those pushed a real language down
That far.
The few tricks that did were extreme, like taking every word and sorting
Its letters into alphabetical order.
Hold on to that alphabetical sorting trick,
Because it matters again in a couple of minutes.
Friedman had seen this shape coming.
His
Study group started on the manuscript in 1944, and he even pulled in the mathematician John von
Neumann.
In 1959, he slipped his conclusion into a footnote about Chaucer, scrambled into
An anagram, with the solution sealed away.
Readers mailed in their own guesses, and the
Journal printed a few, including one that began, this is a trap, not a trot.
When the
Solution was opened after his death, it said he believed the Voynich was an early
Attempt at an artificial universal language, one built from scratch so that the spelling of
Each word files its meaning into categories.
Twelve boxes filled in a fixed order is exactly
What that kind of language would look like.
It’s also what a machine would
Look like.
So after two rounds, the scoreboard stood at one point for a
Real language and one point for a machine.
If a real language is buried under that
Machinery, the next question writes itself.
Which one?
Over the last century,
Serious people have argued for Latin, Greek shorthand, a tonal language from East
Asia, and Nahuatl from Mexico.
In 2019, one paper claimed it was written in something called
Proto Romance, a language no linguist recognizes, and the University of Bristol withdrew
Its own press release after the backlash.
The most interesting attempt came out of the
University of Alberta.
Greg Kondrak and his graduate student Bradley Hauer trained software
On the Universal Declaration of Human Rights, one of the most translated documents on Earth,
In hundreds of languages.
Then they asked which language Voynichese resembled most.
They were
Expecting Arabic, and it came back Hebrew.
So they assumed each Voynich word was a Hebrew word with
The vowels dropped and the letters sorted into alphabetical order, and they wrote an algorithm
To unscramble them.
More than 80 percent of the results turned up in a Hebrew dictionary.
After
A few spelling fixes and a pass through Google Translate, the opening line came out as, she
Made recommendations to the priest, man of the house and me and people.
As opening lines go,
That’s an odd way to start a book full of plants.
Now put that next to the Yale result from a couple
Of minutes ago.
Sorting the letters of each word into alphabetical order was one of the only tricks
That could drag a real language down to Voynich levels of predictability.
And that is exactly the
Scrambling Kondrak’s team built into their test.
Two studies from two different fields, one about
Statistics and one about Hebrew, point at the same hidden mechanism.
It was the most encouraging
Connection the AI found in the entire file.
The Hebrew theory didn’t last long.
A colleague of
The Alberta team who speaks Hebrew natively looked at the raw unscrambled line and said it wasn’t a
Coherent sentence.
Critics pointed out that the method was tilted from the start.
If you drop
The vowels and allow the letters in any order, short Hebrew words built on three consonant roots
Match real words far more easily than English words do.
In English, the word cat only rearranges
Into act.
And Google Translate will turn almost anything into something that sounds like a
Sentence.
So the Hebrew result stays a clue.
But the AI noticed what the clue implies.
If a scrambling procedure can make a real language look like the Voynich, then a
Simpler procedure might make no language at all look like it.
And that reopened
The door the carbon dating had closed.
If you like watching the world’s unreadable
Objects get put through this kind of test, subscribe, because the Voynich is
Only the first file on the table.
The hoax theory had already made one comeback.
In
2004, Gordon Rugg, a computer scientist at Keele University, showed you could churn out Voynich
Style gibberish in a few months using a table of syllables and a Cardan grille, a card with holes
Cut into it that picks pieces out of the table.
By his estimate, a skilled hoaxer could fill most
Pages in an hour or two.
His prime suspect was Kelley.
The trouble is the date.
The Cardan grille
Was invented around 1550, and the calfskin is more than a century older.
Unless somebody was hoarding
Blank vellum for 150 years, Kelley is out.
But the carbon dating only ever ruled
Out a fake made long after the 1430s.
A scribe writing nonsense in 1420 would
Pass every one of those tests.
And in 2019, Torsten Timm and Andreas Schinner described
A way to produce meaningless Voynichese that needs nothing more than a pen and the page in
Front of you.
They call it self citation.
You write a few starter words.
For every new
Word, you glance back at something already on the page and copy it with a small change,
A letter swapped here, a piece added there.
Keep that up for 200 pages and similar words end
Up sitting close together, which is exactly what the manuscript shows.
The text Timm generated
This way also obeys Zipf’s law, the same curve that got treated as proof of language at the
Start.
That’s the nonsense the AI warned about.
Then Yale pushed further.
In 2022,
Daniel Gaskell and Bowern asked 42 volunteers to write short passages of
Invented text by hand.
Human gibberish turned out to be far from random.
It was repetitive and patterned, and in several ways closer to Voynichese
Than real language is, including that strange predictability.
The volunteers kept
Circling back to their own invented words.
So the AI rescored every study on the board.
Every
Oddity at the level of letters and words fits the idea that there’s no message.
Only one piece of
Evidence on the language side is still standing, the topics that line up with the pictures.
The volunteers’ samples were too short to test that.
And a paper Timm released earlier
This year argues that the manuscript’s vocabulary evolves as the book goes on,
Which is what a copying process does, and which could explain why neighboring
Pages share so much vocabulary.
There’s a catch for the hoax theory, and
It cuts the other way.
A cipher clerk who reused recent words to save effort
Would leave the same fingerprints, and researchers on the main Voynich forum
Have been making that argument against Timm for years.
So the AI stopped asking
What the text is, and asked who wrote it.
In 2020, Lisa Fagin Davis, who runs the Medieval
Academy of America, studied the handwriting letter by letter, using digital paleography software
To compare letter shapes page by page, the way a specialist examines any medieval book.
She counted
Five different scribes.
Each one shapes the same few letters with small, consistent habits,
Like how far the tail of one letter curls back, or how many pen strokes go into another.
Scribe
1 wrote the entire opening run of plant pages.
Scribe 2 wrote the second set of plant pages
And the whole bathing section.
On some sheets, the scribes seem to have worked together,
And Davis suspects Scribe 2 may have been running the project.
Not everyone accepts
Her exact count, and Timm is one of the doubters.
But the basic finding, more than
One trained hand, is hard to argue with.
And this is where the second thing the AI found
Turned on the first.
Currier’s two dialects line up with the scribes almost perfectly.
Dialect A is Scribe 1.
Dialect B belongs to the others.
In this book, the subject, the
Dialect, and the handwriting all change at the same places.
So when the vocabulary shifts
Between the plant pages and the bathing pages, you can’t tell whether the topic changed or
The person holding the pen did.
The strongest evidence for meaning might partly be
Evidence for five different writers.
Either way, five people wrote this system
Fluently, with almost nothing crossed out.
That means somebody taught it.
A language
Needs teaching, and so does a procedure.
The people who owned the Voynich manuscript
Afterward tried to learn it too.
The first page carries the faded signature of
Jacobus Sinapius, the emperor’s chemist, recovered under ultraviolet light.
A letter
Found tucked inside the book, written in 1665, claims Rudolf the Second paid 600 ducats for it,
A fortune for a book nobody could read, though nobody has found that payment in the imperial
Accounts.
A Prague alchemist named Georg Baresch spent about 20 years trying to read it and gave
Up only when he died.
He had even copied out pages and mailed them to Athanasius Kircher, a Jesuit
Scholar in Rome who claimed he could read Egyptian hieroglyphs.
It then passed to his friend,
The physician Johannes Marcus Marci.
In 2024, Davis studied multispectral scans of the first
Page and found three faded columns in the margin, a Roman alphabet, a row of Voynich letters,
And a second alphabet shifted by one place.
It’s a key someone was trying to build, and
The handwriting matches Marci’s.
In 1665, he gave up and sent the book to Kircher himself,
And Kircher couldn’t read it either.
After that, the book drifted through Jesuit libraries
And vanished for about 250 years, until it surfaced at the Villa Mondragone and Voynich
Bought it.
Marci sat with this book the way you might be sitting with it right now, lining
Up letters and waiting for one of them to give.
Lay the whole file out at once.
The Zipf curve,
The topic clusters, Currier’s two dialects, the 12 boxes, the line effects, the record low
Predictability, the Hebrew anagrams, the copying method, the volunteers’ fake pages, and the five
Hands.
The AI found one pattern running through all of it, and it’s about distance.
Every test run
On single letters and single words says machine.
Every test run on whole pages and sections says
Language.
And the thing sitting exactly on the boundary between those two scales is the people.
Change the scribe and you change the dialect, the vocabulary, and the subject, all at once.
That’s
Why a century of brilliant work keeps producing opposite answers.
The evidence is tangled,
And the knot runs through five pairs of hands.
The AI also listed what could actually break
The tie, and none of it is out of reach.
Only 10 pages of the manuscript have ever
Been scanned with multispectral imaging, and faded writing already turned up on the
First one, so Davis has been pushing to image every page.
New ink and pigment results from
Over 20 samples were presented at a Voynich conference this March.
And the Yale volunteers’
Experiment needs a sequel at full length, with fake books long enough to test whether
Invented text can produce chapters too.
That one experiment could decide whether the
Last piece of evidence for a language survives.
So this is the verdict the AI would put its name
To.
The Voynich manuscript is a writing system, taught to at least five people and followed with
Almost no mistakes across more than 200 pages, made in the early 1400s on expensive
Calfskin by people who took it seriously.
Whether that system carries a language is
The one question the evidence can’t settle, because in this book the subject, the dialect,
And the writer all change together.
Baresch, Marci, Kircher, and Friedman all assumed
That somewhere, somebody could read it.
The harder question is whether even the five
People who wrote it could have read it back.
I’m Ambrose Wild.
There are more objects out
There like this one, made by people who knew exactly what they were doing and left no
Instructions behind, and we’re putting every one of them through the same cold test.
Stay
Subscribed, because the next file is already open.
