Voynich Manuscript folio 13r, Beinecke MS 408, Yale MS 408 · f13r · Yale Beinecke · Public domain
Short version: a real solution has to do three things. Read coherent, extended text (not a handful of words). Be repeatable, so other people applying the same key get the same reading. And work on pages the author never used. So far, every announced solution fails at least one of these, and most fail all three.

The claims that made headlines

These are the solutions that reached the press and the public. Each was examined by specialists and found wanting.

Claimed Solutions · examined and unconfirmed

Claimed Solution

Gerard Cheshire's "Proto-Romance" (2019)

An academic announced the text was a lost "proto-Romance" language and could be read. The reaction was immediate: linguists pointed out that "proto-Romance" as described is not a recognized language, and that the translation stitched together cherry-picked words from several Romance languages with no consistent grammar. The University of Bristol publicly distanced itself from the paper, and long-time researchers published point-by-point refutations.

Bristol distanced itself · no consistent grammar · refuted 2019
Claimed Solution

Nicholas Gibbs's "Medieval Health Manual" (2017)

In the Times Literary Supplement, Gibbs argued the text was Latin abbreviations describing a women's health manual. Specialists noted that the proposed expansions did not produce coherent Latin, and that much of the reading recycled ideas already considered and set aside. No readable passage followed.

TLS 2017 · no coherent Latin produced
Claimed Solution

Gordon Rugg's "Elaborate Hoax" (2004)

Rugg showed that a Cardan grille, a simple table-and-template trick, can generate text with some Voynich-like statistics, suggesting it could be meaningless. This is a mechanism, not a reading: it shows a hoax is possible, not that the manuscript is one. Later work (Montemurro and Zanette, 2013) found word-clustering patterns consistent with genuine information content, which simple grille gibberish does not reproduce.

a mechanism, not a translation · countered by Montemurro & Zanette 2013
Claimed Solution

Stephen Bax's Partial Decipherment (2014)

A linguist proposed tentative sound-values for a handful of words, mostly plant and star names read as proper nouns. The work was careful and provisional, but it covered only about ten words and was never extended into readable running text. A short, unverified word list is not a solution.

~10 words · never extended to running text
Claimed Solution

"It Is Actually Turkish / Hebrew / Nahuatl"

Every few years someone announces the manuscript is a known language in disguise: Turkish, Hebrew, Manchu, and more. They share a pattern. They fit a few words, declare success, and never produce a coherent, checkable translation of a full page. None has been accepted by Voynich scholars.

recurring · fits fragments, never a full page
Claimed Solution

"AI Cracked the Voynich" Headlines

Since 2018, recurring stories claim a neural network or a chatbot decoded the manuscript. On inspection they produce no verifiable text. A model can always output something, but nothing repeatable, coherent, and checkable against unseen pages has ever appeared. Pattern-matching is not reading.

no verifiable output · not repeatable

Why every solution fails the same test

A genuine decipherment is not a feeling that the words look right. It has to survive three demands.

Test 1

Coherent, extended text

It must read as connected language across paragraphs and pages, not a scattered list of nouns. Most claims stop at a few words.

Test 2

Repeatable by others

Independent scholars applying the same key must arrive at the same reading. Claims that only the author can reproduce are not science.

Test 3

Predicts unseen pages

A real key should read folios the author never used to build it. None of the proposed solutions has passed this.

What we actually know

The honest state of the question is not a translation. It is a set of measurements that any researcher can reproduce.

The text follows Zipf's law and has entropy in the range of real languages. It splits into two statistically distinct hands, the Currier A and B sections, which differ measurably. And on fixed, reproducible structural metrics, the manuscript sits in an intermediate regime: more organized than random gibberish, yet not a clean match to any known language. That gap, not any claimed reading, is the real puzzle.
See our findings → How we measure → What it isn't →
Was this page useful?