The Voynich Manuscript

The manuscript's paper trail is a comedy of people who owned it and couldn't read it either.

The first confirmed owner was Georg Baresch, a 17th-century Prague alchemist who complained that the book — he called it a "Sphynx" — had been "taking up space uselessly in his library" for years. In 1639 he mailed copies to Athanasius Kircher, the Jesuit scholar famous for claiming to have cracked Egyptian hieroglyphs, and asked for help. That 1639 letter is the earliest confirmed mention of the manuscript anywhere. Kircher wanted the book. Baresch wouldn't sell.

After Baresch died, the manuscript passed to Jan Marek Marci, then rector of Charles University in Prague, who finally shipped it to Kircher with a cover letter dated 19 August 1665 or 1666. That letter — still attached when the book resurfaced — carried the story that Emperor Rudolf II had paid 600 ducats for it (about 2.1 kilograms of gold), and that the author might be the 13th-century friar Roger Bacon. The tip came third-hand, from Marci's late friend Dr. Raphael Mnishovsky. Marci was careful:

On this point I suspend judgement; it is your place to define for us what view we should take thereon.

The Bacon attribution has since been largely discredited, but a faint clue points back toward Rudolf's court: the near-invisible signature of Jacobus Horcicky de Tepenecz, head of Rudolf's botanical gardens, still legible under ultraviolet light on the first page. The only matching entry in Rudolf's records is a 1599 purchase of "a couple of remarkable books" from esoteric collector Carl Widemann — plausible, unproven.

The book then vanished into the Collegio Romano library for roughly two centuries, probably filed with the rest of Kircher's correspondence. When Italian troops annexed the Papal States in 1870, faculty quietly shifted books into private libraries to dodge confiscation — which is how the manuscript, still bearing the ex libris of Jesuit rector Petrus Beckx, ended up at the Villa Mondragone near Rome. Wilfrid Voynich, a Polish book dealer, bought it there in 1912 from a discreet Jesuit sale, among thirty manuscripts. It's named after him. After Voynich died in 1930, it moved to his widow Ethel — author of the novel The Gadfly and daughter of mathematician George Boole — then to her friend Anne Nill, then to book dealer Hans P. Kraus, who couldn't find a buyer and donated it to Yale in 1969. It's been Beinecke MS 408 ever since, and in 2020 Yale put the entire thing online.

Here's where the case gets genuinely strange. The text behaves like language — right up until it doesn't.

Roughly 170,000 characters, about 38,000 words with 8,000-odd unique types, written left to right in a script of maybe 20–25 characters with no obvious punctuation. Most characters are just one or two pen strokes. And the writing flows — the ductus is smooth, with no delay between characters, which is not how enciphered text usually looks; encoders hesitate.

The frequency behavior reads like a real language. The most common word appears about twice as often as the second, three times as often as the third. That's Zipf's law — the frequency signature of natural human speech committed to paper. A 2013 São Paulo team led by Diego Amancio ran statistical models and concluded the text was "compatible with natural languages and incompatible with random texts." Word lengths cluster the way real languages cluster: practically nothing under two letters or over ten.

But the internal structure is off in ways no known language does. Certain characters only appear at the start of a word, some only at the end (like the Greek final sigma ς), some only in the middle. Grammatical markers — the equivalents of English -s or -d — never show up mid-word, which Penn State's Gonzalo Rubio flatly called "unheard of for any Indo-European, Hungarian, or Finnish language." Words that differ by a single letter repeat with unusual frequency, and the same common word can appear up to five times in a row.

Then linguists Claire Bowern and Luke Lindemann measured second-order conditional entropy, h2. Natural languages land between 3 and 4. Voynichese comes in around 2 — its character sequences are far more predictable than any human language, yet its higher-level structure looks like one. Bowern's read: not gibberish, but also not consistent with a simple substitution or polyalphabetic cipher.

And it isn't one uniform text. Researchers separate two "dialects": Voynich A dominates the herbal and pharmaceutical pages, Voynich B the balneological and astrological ones, with substantially different common vocabularies. The morphology suggests a three-part word structure — prefix, root, suffix — with variants that seem tied to where a word sits in a line or paragraph. Medievalist Lisa Fagin Davis tracked tiny handwriting tells — loop sizes, crossbar curves, the length of a letter's feet — and concluded the book was written by roughly five scribes, with the variations falling between sections rather than within them. Not one deranged hermit. A community.

People who crack codes for a living have broken themselves on this book.

The roster of failures is elite: William Friedman and Elizebeth Friedman, John Tiltman, Prescott Currier — American and British codebreakers from both World Wars. Friedman ran an informal NSA team on it in the early 1950s and came away convinced the text might be an early attempt at a constructed language, the kind of philosophical classification scheme Bishop Wilkins built in the 1660s. Tiltman, whom Friedman asked to check the pages, didn't buy it — and noted the problem that any such scheme would have been "almost instantly recognisable." In 1962 Elizebeth Friedman flatly called statistical decipherment "doomed to utter frustration." She was, so far, right.

The claimed solutions, by contrast, arrive roughly annually:

• Bacon's microscope diary (Newbold, 1921) — claimed each letter hid microscopic Greek shorthand; John Matthews Manly demolished it in 1931, showing the "markings" were just ink cracking on rough vellum
• Roger Bacon's cipher diary (Feely, 1943) — a substitution cipher for abbreviated Latin; other scholars unanimously rejected the readings
• Vowelless Ukrainian letters (Stojko, 1978) — criticized for loose method and no link between text and images
• A forgery to fool Rudolf II (Brumbaugh, 1978) — Latin behind a two-step cipher
• Hebrew alphagrams (Kondrak, 2017) — computational, first sentence allegedly "She made recommendations to the priest"; the medievalists who reviewed it were not convinced
• Plagiarized women's-health guide in abbreviated Latin (Gibbs, 2017) — Davis said it "doesn't result in Latin that makes sense"
• Phonetic Old Turkic (Ardıç, 2018) — rejected by the journal it was submitted to
• Proto-Romance compiled by Dominican nuns (Cheshire, 2019) — the University of Bristol publicly distanced itself within days

On the other side sits the hoax camp, and it isn't lightweight. In 2003 Gordon Rugg showed a Cardan grille — a perforated overlay of prefixes, stems, and suffixes — could generate Voynich-like text, though the grille itself was invented around 1550, over a century after the vellum. Andreas Schinner's 2007 Cryptologia paper argued the statistics matched quasi-random output. Timm and Schinner's 2019 "self-citation" model reproduced many features by having scribes copy and mutate earlier words. And in 2022, Yale's Daniel Gaskell and Claire Bowern had human volunteers deliberately write meaningless text — the results came out weirdly non-random, with several of the same statistical quirks as Voynichese.

But the meaning camp has its own data. Marcelo Montemurro's 2013 information-theory study found content-bearing words clustering the way they do in real books, with new vocabulary appearing at topic shifts — patterns he argued nobody in the 1400s had the academic tools to fake on purpose. A 2021 Yale tf–idf study found computational topic clusters matching the illustrated sections. Lisa Fagin Davis's verdict on the would-be translators is the one that sticks: their logic is "circular and aspirational," the translations "themselves aspirational rather than being actual translations."

Two things are settled. The rest isn't.

Settled: it's not a modern forgery. The radiocarbon date and the fourteen-plus calfskins of unused early-15th-century parchment make that, in Eamon Duffy's word, "inconceivable" — you couldn't fake it with material a forger like Voynich could plausibly source. The June 1999 discovery of Baresch's letter, which places the book in Prague long before Voynich was born, closed the fabrication case for most researchers. Settled: it has never been demonstrably deciphered, and none of the proposed solutions have been independently verified.

What's left is the uncomfortable middle. The text is too organized to be pure noise — Zipf's law, sectional vocabulary, multiple consistent scribes — and too weird for any cipher anyone has tested, with that stubbornly low h2 entropy. A 2025 Cryptologia paper proposed a "Naibbe" verbose substitution cipher that reproduces many of the manuscript's statistics using only 15th-century tools, building on the observation that tight little repeating shapes like "or or oro r" look a lot like Roman numerals run through a verbose encoder. The author was explicit: proof of concept, not the solution. That's the honest posture the whole field should probably borrow.

The author is anonymous, the origin possibly Italy, the date early 1400s — and no named candidate has survived scrutiny. Voynich himself, Roger Bacon, John Dee, the Florentine architect Antonio Averlino: all proposed, all unconfirmed.

The purpose stays murky too. The soft, heavily-thumbed parchment says practical reference work — a medical manual, a celestial almanac — the kind of book you page through, not the kind you enshrine. But not one plant drawing is unambiguously identifiable, and several appear to be composites: roots of one species fastened to the leaves of another, with flowers from a third. Only a wild pansy and a maidenhair fern can be named with any confidence. In September 2024, Yale released multispectral scans of ten pages, surfacing details invisible under normal light. That analysis is ongoing.

The puzzle is clean when you strip it down. A book worn soft by handling. Text that obeys Zipf's law but breaks every rule of word-internal grammar. Five hands writing one consistent unknown script. More than a century of the best cryptanalysts on Earth — and still, nobody can read a word of it.