The Voynich manuscript is an undeciphered book hiding secret knowledge

The steelman: this really does look like language

Set aside the sensational “solved!” headlines, because underneath them sits a genuinely serious case, made by working scientists and linguists rather than hobbyists, that the Voynich manuscript is exactly what it looks like: a real book carrying real content. Start with the object itself. The vellum was radiocarbon-dated at the University of Arizona in 2009 to 1404–1438, with separate microscopy of its inks and pigments finding materials consistent with that same period: a genuine fifteenth-century artifact, not a later curiosity dressed up to look old. Whoever made it invested serious, specialist effort across several technical subjects at once: roughly 113 careful botanical drawings, a run of astronomical and astrological star charts and zodiac roundels, and the section of bathing figures described above (sometimes labeled “balneological,” after the medieval genre of writing about the therapeutic use of baths) showing dozens of women moving through interconnected pools and pipework in what reads as a deliberate procedural sequence, not idle decoration. That is a great deal of coordinated, confidently executed work to produce if there was nothing organizing it underneath.

The most basic textual test is one anyone can check without specialist training: the manuscript's word frequencies follow Zipf's law, the same rank-frequency pattern found in every natural language ever measured, where the single most common word appears roughly twice as often as the second most common and three times as often as the third. That is not the kind of curve idle, meaningless scribbling produces by accident; whatever generated this text was following rules regular enough to reproduce one of the basic statistical signatures of real language. In a peer-reviewed 2013 study published in PLOS ONE, physicists Marcelo Montemurro and Damián Zanette pushed further with information-theoretic methods (the same kind of statistical tools used to study word distribution in known languages) and found that the manuscript's words cluster thematically the way real content words do: certain terms concentrate heavily in the botanical section, others in the biological section, in patterns that track the book's own illustrated subject matter. Random gibberish, generated without any underlying structure, does not organize its “keywords” by topic like that. The authors concluded that this “limits severely the scope of the hoax hypothesis” and lends support to “the presence of a genuine message inside the book.”

That finding sits alongside decades of work by linguists including Claire Bowern and Luke Lindemann, whose 2021 survey in the Annual Review of Linguistics notes that measures such as word-length distribution, character co-occurrence and second-order entropy (a statistic capturing how predictable one letter is given the one before it) place Voynichese in a range that overlaps with real historical languages more than with purely random text. If anything, Voynichese comes in at lower entropy than most real languages at the character level (a second-order entropy near 2, against roughly 3 to 4 for typical natural languages), meaning each letter is, if anything, more predictable from the one before it than in ordinary text, the opposite of what meaningless noise should look like. A separate 2011 study by computational linguists Sravana Reddy and Kevin Knight, presented at an Association for Computational Linguistics workshop, compared Voynichese's letter statistics directly against real writing systems and found the closest match was an abjad, a consonant-centered script structure used for languages like Arabic and Hebrew, a specific, checkable resemblance to how an actual historical writing system is built, not merely to language statistics in the abstract. This is not a fringe finding smuggled past reviewers: multiple independent teams, using different statistical toolkits across different fields, have converged on compatible conclusions, and none of them is in the business of promoting fringe decipherments.

Then there is the resistance itself. This is not a document that a hobbyist glanced at and gave up on. William F. Friedman, arguably the most accomplished American cryptanalyst of the twentieth century (a man who helped break Japan's wartime diplomatic cipher and went on to become the NSA's first chief cryptologist), spent years of off-hours attention on the Voynich manuscript, at one point organizing an informal study group of fellow codebreakers and mechanizing the transcription onto IBM punch cards to search for patterns by machine. He did not work alone: his wife Elizebeth Smith Friedman, a pioneering cryptanalyst in her own right who had introduced William to the field before building a celebrated career of her own breaking Prohibition-era smugglers' codes and a Japanese espionage ring's cipher, joined the study directly. By 1962 she stated in print that simple single-substitution-alphabet decipherings were “doomed to utter frustration”: the judgment of a specialist who had spent years on the text, not a dismissal of the manuscript itself, that whatever it is, it was not going to yield to the simplest and most obvious kind of code. Other NSA analysts picked it up independently. None of them broke it. A text that defeats sustained attention from people whose entire careers were spent breaking other people's secrets is not, on its face, an obviously trivial hoax; it is at minimum a very well-constructed one, and the honest reading of the entropy and keyword evidence is that “well-constructed hoax” and “real but unrecognized language” are still both on the table.

Why every proposed solution has collapsed

Statistical structure is not the same thing as a translation, and on the specific claim that anyone has recovered actual meaning from the manuscript, the record is one of unbroken failure. Over more than a century, professional and amateur researchers have proposed the text is: Latin, Hebrew, Ukrainian, a Chinese-derived script, a lost dialect of proto-Romance, an artificial or constructed language, a substitution cipher, a verbose cipher hiding a much shorter plaintext, and outright meaningless glossolalia produced by a table-and-grille hoax generator. Each proposal has, so far, either failed to produce a translation that other specialists can independently reproduce from the same method, or produced “translations” that read as incoherent or grammatically broken once checked by an actual expert in the claimed source language.

The two most publicized modern cases illustrate the pattern. In 2014, applied linguist Stephen Bax published a proposed partial reading identifying roughly ten words and fourteen characters by matching plant and star illustrations to proto-Romance or Semitic root words. Fellow researchers, including cryptographer Gordon Rugg, found the method unconvincing: a handful of speculative word matches, drawn from a huge space of possible readings, is not strong evidence on its own, and no further words were successfully added to the list using his method before Bax's death in 2017. In 2017, television writer and researcher Nicholas Gibbs announced in the Times Literary Supplement that the manuscript was a women's health manual written in abbreviated medieval Latin. Medievalists moved quickly: Lisa Fagin Davis, executive director of the Medieval Academy of America, and others pointed out that Gibbs's sample “translations” did not parse as grammatical Latin at all, and that his method could be used to generate almost any desired reading from the same symbols. Both claims generated a wave of credulous news coverage before the correction caught up, a pattern that has repeated with nearly every subsequent “cracked it” announcement, including AI-assisted attempts in the 2020s that produced fluent- sounding but unverifiable output.

The statistical arguments deserve the same scrutiny as the specific translation claims, and they hold up only partway. Voynichese does follow a Zipf's-law-like word-frequency curve and shows unusually low character-level entropy, but that same Gordon Rugg, in a 2016 study in Cryptologia co-authored with Gavin Taylor, found that the table-and-grille hoax method already noted above, used to generate batches of admittedly meaningless gibberish, reproduced Zipf's law and the manuscript's other unusual quantitative quirks just as well. A statistical signature that both real language and a specific, documented hoax method can produce is not a signature that decides the question either way. The 2011 comparison to an abjad script fares similarly: its own authors noted that the resemblance does not by itself explain Voynichese's unusually low conditional entropy, leaving the text's core predictability puzzle open even within their own preferred framework. Nor does the Friedmans' own verdict settle it: their conclusion, reached after decades of failed cryptanalysis, was that the manuscript most likely represented an invented or constructed language rather than a straightforward cipher over a real one, a judgment that rules out the simplest “secret message” theories but says nothing about whether that invented system was meant to communicate anything at all, or was built only to look like it did.

René Zandbergen, among the most active specialists tracking the manuscript's scholarship for decades, has repeatedly made the same point to journalists: whatever headline you just read, the text of the Voynich manuscript has not been solved. No decipherment claim to date has met the basic bar of scholarly acceptance: reproducibility by independent researchers, internal grammatical consistency, and translated content that makes sense against the manuscript's own illustrations rather than being fitted to them after the fact.

An authentic mystery, and the perfect surface to project onto

Unlike many entries in this encyclopedia, the Voynich manuscript is not driven by distrust of institutions or a conviction that officials are lying: the Beinecke Library has done the opposite of covering it up, digitizing the entire book and publishing it freely online precisely so anyone in the world can study it. What sustains belief here is something rarer and in some ways more interesting: a genuine, professionally certified unsolved problem, sitting in plain public view, with just enough real statistical structure in the text to make “it must mean something” a reasonable-sounding instinct rather than an obviously paranoid one.

The illustrations do real psychological work too. Pages of unidentifiable plants, nude figures bathing in green liquid through networks of tubes, and star charts with no confirmed astronomical referent invite exactly the kind of pattern-completion that esoteric and lost-civilization narratives thrive on: the pull to read alchemy, forgotten medicine, or hidden cosmology into images that resist any confirmed identification. And because a real, peer-reviewed finding exists (that the text's statistics resemble natural language more than randomness), it is easy, and only slightly wrong, to compress that into “scientists proved it's a real language,” skipping the far less satisfying actual conclusion: the structure is consistent with meaning, and also consistent with an unusually well-built hoax, and no one has shown which.

Finally, there is the prize on offer. As one of the last major undeciphered historical documents (freely available in full to anyone with an internet connection, unlike a classified file or a hidden artifact) it functions as an open scholarly competition with an enormous reward: whoever produces a decipherment that actually survives scrutiny would join the ranks of history's celebrated codebreakers. That incentive, combined with how easy it is to generate a plausible-looking partial match in a script no one can check against a known language, guarantees a steady stream of new claimants, and, so far, a perfect record of those claims not holding up.