THE VOYNICH MANUSCRIPT SHOWS LITTLE TRACE OF LANGUAGE

The Voynich manuscript is a book of about 240 pages that radiocarbon dating places in the early fifteenth century. Drawings of plants, stars and bathing figures surround some 38,000 words in a script found nowhere else. No one can read it.

It looks like writing: words separated by spaces, many rare words, signs — called glyphs — that keep to fixed places inside words. Yet it behaves like no known language. The same word often appears two or three times in a row, many words differ from their neighbours by a single glyph, and the frequent words are short and alike.

Gibberish, or a language in disguise?

In one account, the text means nothing: a scribe wrote each new word by copying an earlier one and changing a glyph or two, a process proposed by Timm and Schinner in 2020. In the other, a real language hides behind a cipher — one recent design, the Naibbe cipher, turns Latin or Italian into text much like the manuscript’s.

Each side has long argued from statistics chosen for the purpose. Xin Zhang and Victor S. Sheng, at Texas Tech University, judged both under the same rules.

Same pages, same budget, unseen statistics

Each account became a program that fills the manuscript’s pages and lines, word by word. Copying programs write each word as new or as a changed copy of an earlier one. Cipher programs encrypt a real medieval text, replacing each word with a code word, or its letters, its syllables or only its grammatical class. Every cipher also gets a copying scribe, so that local repetition favours neither side.

Each program was tuned on 22 statistics of the text, then judged on 15 others it never saw.

Timm and Schinner’s copying rules predict those unseen statistics far better than any cipher with a single key. Their worst miss stays under five standard deviations (the typical spread of the program’s own texts); a word code misses one by more than a hundred.
The decisive statistic is a drift. The manuscript has two “dialects”, Currier A and B, which use glyphs differently. A fixed key writes the same glyph habits on every page; a scribe who builds words from words he has just written lets them wander.
But ciphers with a second key — a separate key for the Currier B pages — or that encode only the word classes of a real text, do just as well, with a scribe who miscopies about half the words. In the authors’ words, “the ciphers that fit best are those whose plaintext matters least” (the plaintext is the original text, before encryption).

Word statistics alone cannot decide, and no program matches all 37 statistics at once.

Words that keep meeting on a line

So the authors looked for a trace of language that needs no decipherment. In a meaningful text, words that belong together — a noun and its verb, the parts of a set phrase — share a line more often than chance, and do so again elsewhere in the text. A word code keeps this pattern; a pair born from copying on one page has no reason to return pages later.

The test splits the book in two and checks whether the same pairs of frequent words keep sharing lines in both halves. Identical words and words one glyph apart are left out, so copying cannot fake the signal.

28 texts in 16 languages, laid out on the manuscript’s pages, score between 0.139 and 0.411 (a correlation between the two halves: 0 when the pairs found in one half say nothing about the other, 1 when they match perfectly).
The manuscript scores 0.071, and 0.053 in a second, independent transliteration (a typing of the glyphs into a standard alphabet): half the strength of the weakest text.
Word codes sink to that level only when their original text shows the pattern weakly — above all medieval Latin written with its abbreviations, at 0.102 — or when many words are miscopied.

Calibration matters: a program trained to tell volunteers’ gibberish from real prose calls the manuscript gibberish, but it also calls 53 of 54 enciphered meaningful texts gibberish.

Code-breakers that must read known texts first

Each solver searches for the key under which the text reads most like a given language, scored with a simple statistical table of which letter, or word, tends to follow which in old texts of that language. Each was first run on control texts, real texts enciphered the same way; where it could not read them, the test counted as undecided.

Letter ciphers. The solver read the controls of 15 of 16 languages. In each, the manuscript showed less order than the controls, and its readings were strings of letters without recognisable words. Hebrew stayed undecided.
Word codes. For a code of abbreviated Latin, the original text the cipher programs preferred, a table built from 5 million words of such Latin still found more word order in the controls than in the manuscript, unless two words in five were miscopied.
One word per letter or syllable. 28 such schemes gave 1,144 decisions: 596 excluded, 429 undecided, and 119 that could not be excluded: the manuscript edged just above texts with no message at all, yet decoded into nothing readable.

The cipher that slips through

A nomenclator, a kind of code common in the period, writes the most frequent words as whole code words and spells the others letter by letter. With a second key and a scribe who miscopies a large share of the words (37% and 56% in its two best fits), it matches the manuscript’s word pairs, and the solvers cannot exclude it in 9 of the 16 languages. A cipher whose key changes along the text was not tested at all.

A seam between words

The strongest structure left sits at word boundaries. The last glyph of a word constrains the first glyph of the next: 0.204 bits of mutual information (how much one tells you about the other), against 0.078 for the copying program and 0.014 for the ciphers. This link carries most of what the letter solver found. The authors trace it to glyphs bound to their neighbours twice as tightly as letters in the languages tested; no program reproduces it together with the manuscript’s spaces.

Room for a short message

The authors also planted messages of 50 to 5,000 letters in the manuscript’s text and searched for them. Only one kind is reliably caught: each letter always written as the same word, the words side by side. Hidden in the first words of paragraphs, every such message of 500 letters was found. Nothing in the manuscript points to one — but a message using several words per letter, or with its letters spaced apart, could still hide there.