Voynich Manuscript — Theories and Decipherment Attempts

Six hundred years of decoders have failed on a 240-page book whose vellum is dated 1404–1438, whose statistics look like a real language, and whose every claimed translation has collapsed under scrutiny. The interesting development of 2025 is not a decipherment. It is a working proof that a 15th-century Italian scribe with a deck of cards and a die could have produced text statistically indistinguishable from Voynichese. See concept voynich manuscript for the artifact itself.

As of May 2026, no decipherment has been accepted by the scholarly community.

The current standing of theories

Theory	Standing	Main proponents	Fatal objection
Verbose homophonic cipher (Latin/Italian)	Mainstream, viable	Greshko (2025), Naibbe paper	No decipherment produced
Hoax / structured gibberish	Mainstream, gaining ground	Rugg (2004), Schinner (2007), 2025 volunteer study	Production cost; five-scribe collaboration
Constructed language	Plausible minority	Zandbergen	Statistics should look more artificial
Unknown natural language	Minority	Various	Should have yielded by now under modern methods
Hebrew (AI-claimed)	Fringe	Hauer & Kondrak (2019)	Method removed vowels; no coherent passage
Old Turkish	Fringe	Altounian et al. (2022)	No accepted reconstruction
Alien / magical / horoscope	Fringe	Popular media	No serious support

The Naibbe cipher — 2025's most important development

Greshko's paper in Cryptologia (November 2025) is named for naibbe, the 14th-century Italian card game ancestral to tarot. It proposes a verbose homophonic substitution cipher: each plaintext letter maps to multiple Voynich glyph strings, with padding glyphs inserted to control word length. The encoder rolls a die to chunk the plaintext, draws a playing card to select one of six substitution tables, and writes out the result. All required tools existed in 15th-century Europe. The procedure runs entirely by hand.

What it reproduces from Voynichese:

Symbol frequency distributions
Word length distributions
Positional constraints (certain glyphs only at word start or end)
Entropy levels
Zipfian word frequency

What it does not do:

Decode the manuscript
Prove this was the actual cipher used
Explain the illustrations

The point is the existence proof. Skeptics had argued for decades that no plausible historical cipher could yield Voynichese statistics. That argument is now closed. The cipher hypothesis is mechanically viable. Greshko stresses this is a proof of concept, not a claimed solution — many cipher systems can reproduce statistical signatures.

The 2024 multispectral discovery

In September 2024, Lisa Fagin Davis and Roger Easton reprocessed multispectral images of folio 1r from the Lazarus Project's 2014 capture. Three columns of letters, invisible to the naked eye, emerged: Roman alphabet A–Z, Voynich characters, Roman alphabet offset by one letter. The offset between the two Roman columns is a single-shift Caesar cipher — the simplest possible substitution.

Fagin Davis read this as an early owner's failed decoding attempt, likely 17th century, consistent with the Marci-Baresch correspondence in which the manuscript was sent to Athanasius Kircher for solving. Two things follow. First, contemporaries also believed it was a substitution cipher. Second, the simple cipher did not work — the actual mechanism, if there is one, is more complex.

The gynecology hypothesis (Brewer and Lewis, 2024)

Keagan Brewer and Michelle Lewis, writing in Social History of Medicine (Oxford, 2024), argue the manuscript encodes gynecological and reproductive medicine — contraception, abortion, conception — using cipher because contemporaneous practitioners explicitly recommended secrecy. Their strongest piece of period evidence is Johannes Hartlieb (c. 1410–68), a Bavarian physician contemporary with the manuscript's vellum, who advised using "secret letters" to conceal medical texts on these subjects.

They reinterpret the nine-panel Rosette fold-out as a coitus-and-conception diagram, with anatomical details in the upper-left panel allegedly corresponding to al-Rāzī's description of female anatomy. The balneological section, traditionally read as women bathing for leisure, becomes a gynecological reference plate.

The hypothesis shifts the question from "what language" to "what subject." If the topic is established independently, it constrains candidate cipher types and vocabulary. The interpretive latitude required is the main objection — no decoded passage corroborates the reading, and alternative interpretations of the Rosette page remain plausible.

Why AI has not solved it

Hauer and Kondrak's 2019 paper (University of Alberta) claimed neural machine translation identified Hebrew as the most likely source, with 80%+ of Voynich words matching Hebrew dictionary entries. The method removed vowels and treated words as consonant skeletons — a matching rule loose enough that comparable hit rates can be obtained for many texts in many languages. No coherent Hebrew passage was produced.

A 2021 Yale-affiliated group applied LDA, LSA, and NMF topic modeling and found genuine vocabulary clusters across sections — herbal vocabulary is structurally different from balneological vocabulary. This supports intentional organization but does not identify content.

GPT-class models, applied informally and semi-formally from 2024 through 2026, have produced no accepted decipherment. The reason is structural. Without external ground truth — a known translation of even a few words — large language models cannot distinguish the correct interpretation from infinitely many wrong ones that satisfy the same statistical constraints. The Voynich is a case where the bottleneck is pattern validation, not pattern recognition. More compute does not help.

The gibberish case is stronger than it used to be

A 2025 volunteer experiment (reported by The Art Newspaper) asked participants to produce pages of "convincing gibberish." The results showed Zipfian word frequencies, naturalistic word-length distributions, and several statistical properties once cited as proof of real linguistic content. Humans are surprisingly good at generating statistically language-like nonsense.

This does not prove the Voynich is gibberish. It weakens the statistical evidence that ruled gibberish out. Combined with Rugg's 2004 table-and-grille mechanism and Schinner's 2007 stochastic analysis, the hoax hypothesis is no longer the joke position.

The Currier A / Currier B problem

Any theory must explain this: the manuscript contains two statistically distinct text types, written by five distinct scribes per Fagin Davis's paleographic analysis. Currier A appears in herbal and pharmaceutical sections, Currier B in balneological and astrological sections.

If cipher: two ciphers, or one cipher applied to two source languages
If hoax: collaborative, requiring multiple people to maintain consistent statistics
If constructed language: two dialects, or two stages of construction
If natural language: two source languages, two regional dialects, or stitched-together exemplars

The five-scribe finding is the harder constraint. It cuts against the "lone eccentric" hoax and points toward institutional production — a scriptorium or a group with shared method.

What's contested

The deepest contested question is whether the manuscript carries semantic content at all. The 2025 Naibbe paper makes cipher mechanically viable; the 2025 gibberish study makes hoax statistically viable. Both cannot be the answer. Neither has been ruled out. The scholarly community is split roughly between "complex cipher of a real language, not yet broken" and "structured nonsense produced by a method we have not yet identified."

A second contested question: what would count as a solution? Scholars converge on four criteria — decode at least one full coherent passage; the passage must be independently verifiable against a known historical text or formula; the key must work across multiple unconnected passages; and ideally, external corroboration of the cipher system. No proposed decipherment has cleared all four. Most have cleared none.

The fringe theories are closed cases. Roger Bacon authorship is ruled out by radiocarbon dating — he died in 1292, the vellum is 1404–1438. John Dee forgery is ruled out by the same dating. The Mesoamerican-botanicals theory dies on the vellum predating Columbus.

Why this has to do with other realms

The Voynich is the cleanest case study in the limits of statistical inference. Modern cryptanalysis, modern linguistics, and modern AI converge on the same wall: without anchored ground truth, statistics alone cannot decide between a real language under unknown encoding and structured nonsense generated by an unknown method. This is the same epistemological limit that haunts concept fermi paradox — observing a signal does not tell you whether it carries meaning, and the absence of confirmable decoding does not let you conclude there is nothing to decode. Pattern recognition is cheap. Pattern validation requires an external anchor the universe has not provided.

An open question

The 2026 International Conference on the Voynich Manuscript, organized under the Medieval Academy of America, is the first large interdisciplinary gathering specifically targeting the manuscript. If a serious resolution is going to emerge in this decade, the proceedings will indicate where. If consensus instead lands on "structured gibberish, mechanism unknown," that is a different and possibly more humbling result. Which way will the field move when the cipher and hoax hypotheses both have working mechanical proofs and neither has external corroboration?

Key sources

Greshko, "The Naibbe Cipher" — Cryptologia, November 2025. The verbose homophonic cipher proof-of-concept; code on GitHub.
Brewer & Lewis, Social History of Medicine, Oxford, 2024 — the gynecology and Hartlieb-cipher-recommendation hypothesis.
Fagin Davis, Lisa — paleographic analysis identifying five scribes; 2024 multispectral folio 1r findings (with Roger Easton, Rochester Institute of Technology).
Rugg, Gordon (2004) — table-and-grille hoax mechanism; Cryptologia.
Schinner, Andreas (2007) — stochastic analysis supporting hoax viability; Cryptologia.
Hauer & Kondrak (2019), University of Alberta — the Hebrew/NMT claim, primarily useful now as a cautionary methodological case.
To verify: full proceedings of the 2026 International Conference on the Voynich Manuscript (Medieval Academy of America).

Abhishek