A Century of Voynich Research: From Newbold to AI

The Voynich Manuscript has not spent the last century waiting passively for one genius to arrive.

The manuscript stayed where it was.

The questions moved.

At first, the great question was: Who wrote this?

Then: What cipher is this?

Then: Does the text contain more than one statistical variety?

Then: How do we transcribe an alphabet we do not understand?

Then: What can statistics tell us before meaning?

Then: How old is the parchment?

Then: How many scribes?

Then: Can computation distinguish language, cipher, notation and generation?

And now, inevitably: Can AI solve it?

The important history is not a procession of people who failed. It is a history of instruments. Each generation acquired a new way to see the same book.

The deepest progress in Voynich research often came not from a new answer, but from a better representation of the question.

Quick Read

Seventeenth-century owners such as Baresch and Marci already treated the manuscript as difficult and sought expert help from Athanasius Kircher.
After Wilfrid Voynich acquired the codex in 1912, Roger Bacon became the dominant early authorship hypothesis.
William Romaine Newbold proposed an elaborate decipherment tied to Bacon; John Matthews Manly’s criticism became an early lesson in overfitting pattern-rich material.
Theodore Petersen spent years making a painstaking hand copy and concordance, creating research infrastructure rather than a grand solution.
William F. Friedman brought professional cryptanalytic group work to the manuscript in the 1940s.
John Tiltman stressed that ordinary classical cipher assumptions did not explain Voynich comfortably.
Prescott Currier’s 1976 presentation showed that the manuscript contains at least two strong textual varieties, now called Currier A and B.
Mary D’Imperio synthesised cryptanalytic, statistical, palaeographic and historical work in An Elegant Enigma and related research.
Digital transliteration turned the manuscript into a corpus that could be searched, counted and shared.
Modern linguistics, information theory and statistical modelling revealed non-random structure without producing an accepted translation.
Radiocarbon dating, materials analysis and multispectral imaging transformed the codex into a measurable physical object as well as a cryptographic puzzle.
Digital palaeography reopened questions of scribal hands and production organisation.
Modern AI can accelerate transcription, clustering, comparison and hypothesis testing, but pattern discovery remains different from reproducible decipherment.

Before 1912: The Manuscript Was Already a Problem

The story does not begin with Wilfrid Voynich. By the seventeenth century, Georg Baresch in Prague possessed a manuscript he could not read. He sent copied material and questions toward Athanasius Kircher in Rome, whose reputation for languages and ancient scripts made him an obvious candidate for help.

Later, Johannes Marcus Marci sent the actual manuscript to Kircher with the famous covering letter that preserves the Rudolf II and Roger Bacon stories.

That early history gives us one important control: Voynichese was not made mysterious by twentieth-century ignorance. Learned readers much closer to the manuscript’s early life already found it opaque.

1912: Wilfrid Voynich Changes the Scale of the Mystery

Wilfrid Voynich acquired the manuscript in 1912 from Jesuit holdings at Villa Mondragone near Rome. His great contribution was not decipherment. It was mobilisation.

He recognised that the object could attract scholars, historians, cryptographers and collectors. Voynich became strongly invested in Roger Bacon, and the Bacon hypothesis shaped the first major modern research era because it supplied a famous author before the text had supplied a readable sentence.

An attractive identity can organise research powerfully even when the identity itself later fails.

Newbold: When More Magnification Created More Meaning

William Romaine Newbold produced one of the most famous early Voynich solutions. He believed microscopic details inside the visible characters encoded further information and used elaborate interpretive transformations to recover a text associated with Roger Bacon.

The theory was ingenious. It was also a warning about representation. If a method is permitted to treat accidental ink irregularities as intentional micro-signs, a manuscript becomes extraordinarily rich in apparent hidden structure.

John Matthews Manly’s later criticism showed how unstable such readings could be. The useful lesson is not that early researchers were foolish. It is that increased pattern resolution can manufacture false information when the representation itself is not validated.

Theodore Petersen: A Different Kind of Achievement

Theodore Petersen’s work is less famous outside specialist circles and arguably more important to the long-term research culture. Beginning in the 1930s, he undertook the exhausting task of making a detailed hand copy of the entire manuscript, consulting the original for difficult passages.

He completed the copy in 1944. He annotated pages, marked unusual sequences, highlighted frequent words, compiled plant identifications from collaborators and created a word concordance.

Before a mystery can be analysed at scale, somebody has to turn it into data.

A hand transcription is not glamorous. It is infrastructure.

Friedman: The Puzzle Enters Professional Cryptanalysis

William F. Friedman brought a new institutional culture to Voynich research. He was one of the leading cryptanalysts of the twentieth century, and in the 1940s he organised a study group to work systematically on the manuscript.

The group created transliteration conventions, tabulations and statistical material that allowed researchers to compare observations rather than work only from visual intuition.

This changed the question again. Instead of asking which famous medieval genius wrote the book, researchers could ask what signs exist, how often they occur, which combinations are common and which cipher families fit or fail.

Friedman later considered artificial or synthetic-language possibilities rather than insisting on one ordinary classical cipher. That willingness to change mechanism family was itself a research advance.

Tiltman: Stop Expecting an Ordinary Cipher

John Tiltman, another major cryptanalytic figure, examined Voynichese and reached an important negative conclusion: the manuscript did not behave comfortably like a straightforward classical substitution cipher.

This does not mean “not a cipher”. It means the easy cipher family was not enough.

Tiltman also considered synthetic-language possibilities and paid close attention to internal word structure. Once again, progress arrived by shrinking an overbroad possibility space.

Currier: The Manuscript Stops Being One Text

Prescott Currier’s 1976 presentation changed the manuscript conceptually. He showed that Voynichese is not statistically uniform across the codex.

Currier identified two strong textual varieties that came to be called Currier A and Currier B, along with observations about scribal hands.

This destroyed a hidden assumption. Researchers could no longer safely treat every Voynich page as one homogeneous sample.

Before asking what Voynichese means, Currier forced researchers to ask how many Voynichese-like states the manuscript contains.

The dedicated Currier A and B article owns that discovery in detail.

Mary D’Imperio: The Field Gets a Synthesis

Mary D’Imperio played a central role in turning scattered cryptanalytic work into a research literature. Her work brought together historical theories, transcription systems, statistical results, scribal questions and competing explanations accumulated over decades.

Her 1978 The Voynich Manuscript: An Elegant Enigma became one of the field’s foundational syntheses. The manuscript was no longer merely a code to crack. It had become a multi-disciplinary research object with an accumulated history of constraints.

The Transliteration Revolution: Turning Ink Into a Corpus

Computers need symbols. Voynich gives them handwriting.

That mismatch produced one of the most consequential modern research efforts: standardised transliteration.

Earlier systems—including Friedman-group and Currier transliterations—allowed characters to be represented consistently enough for counting. During the later twentieth century and early internet era, researchers worked toward interoperable conventions that could be shared across software and research groups.

EVA—the European Voynich Alphabet—became a widely used convention for representing visible glyph shapes with ordinary keyboard characters.

This was a huge advance and a dangerous convenience if misunderstood. EVA q does not mean the sound /q/. EVA d is not the decoded Latin letter d. The characters are labels for visible forms.

Transliteration made large-scale analysis possible precisely by postponing translation.

See EVA, Transcription and the Segmentation Problem.

The Internet Changes Who Can Work on Voynich

Before high-quality digital images and public transcriptions, serious Voynich work required access to rare reproductions, photostats, microfilm or the manuscript itself.

The internet changed the research population. A linguist could download text. A programmer could count tokens. A historian could compare folios with digitised manuscripts elsewhere. An amateur could inspect a page at high resolution.

This democratisation was enormously productive. It also produced an explosion of weak solutions because access to data is not the same as access to methodological discipline.

Voynich entered the age of mass hypothesis production.

Statistics: The Question Becomes “What Does the Text Do?”

Once Voynichese became a machine-readable corpus, researchers could study behaviour without deciding meaning first.

frequency distributions;
word lengths;
neighbour relationships;
entropy;
vocabulary growth;
topic-like clustering;
line-position effects;
local repetition.

This phase created one of the most durable shifts in Voynich research: meaning is not the only thing a text can reveal. A system can be described structurally before it is deciphered.

The danger moved too. A statistic can be real while the explanation attached to it is wrong. This is why the modern research estate needs separate owners for predictability, word families, syntax, repetition, local vocabulary and frequency laws.

See Why Voynichese Is So Predictable, Word Families, Syntax Before Semantics and Frequency Laws.

Modern Linguistics Reframes “Looks Like Language”

Claire Bowern and Luke Lindemann’s 2021 review brought modern linguistic methods into a careful synthesis of Voynich research.

The important contribution was not a translation. It was a framework for asking linguistic questions without pretending the language had been identified.

What properties resemble phonology?
What resembles morphology?
Which distributions resemble natural-language structure?
Which proposed language families survive comparison?

This helped replace the vague sentence “Voynich looks linguistic” with measurable claims. The boundary remained essential: language-like is not language-identified.

Material Science Changes the Question Again

For much of modern research, the manuscript was treated primarily as an intellectual puzzle. Scientific testing forced attention back to matter.

Radiocarbon dating constrained the parchment to an early-fifteenth-century material horizon.
Materials analysis examined inks and pigments.
Protein work identified calf parchment.
Multispectral imaging revealed faint and otherwise difficult marks.

These methods did not decipher one sentence. They transformed the permissible history.

Roger Bacon as direct author disappeared. A simple modern fabrication became untenable. Production layers became measurable. The manuscript acquired a much more precise material horizon.

See Ink, Pigment, Parchment and the Physical Evidence.

Multispectral Imaging: The Page Becomes More Than Visible Light

Modern imaging makes another historical leap possible. A researcher no longer has to accept what ordinary visible light reveals.

Different wavelengths can enhance faded marks, overwritten material, offset and pigment differences. This is especially important for later marginal writing, ownership traces and subtle production evidence.

The instrument changes the question from “What can I see?” to “Which physical layer becomes visible under which conditions, and when was it added?”

Digital Palaeography Reopens the Scribe Problem

Handwriting used to be compared almost entirely by expert eye. High-resolution digital images allow researchers to compare glyph construction across the whole codex at scale.

Lisa Fagin Davis’s influential work proposed five distinct scribal hands. Later criticism has questioned whether the variation is best represented by five discrete people or by a more continuous pattern.

The advance is not merely the number five. It is the move from “this page looks different” to “which repeatable palaeographic features cluster across hundreds of pages?”

See The Scribe Problem.

The Mechanism Debate: Language, Cipher, Generation or Hybrid?

One of the healthiest recent changes is that research increasingly separates properties from mechanisms.

Low conditional entropy is a property. A cipher is a mechanism.
Word families are a property. Local generation is a mechanism.
Language-like frequency is a property. Natural language is a mechanism family.

Modern studies can therefore ask whether a proposed mechanism reproduces several independent properties simultaneously.

The 2025 Naibbe work is useful in exactly this way: it demonstrates that a historically plausible hand-operable cipher can produce many Voynich-like statistical properties while preserving plaintext. That does not solve Voynich. It changes which cipher possibilities can be dismissed merely from selected statistics.

See Cipher, Plaintext or Generated System?.

Then Comes AI

AI changes the scale of pattern search.

cluster pages;
compare glyph variants;
suggest transcription candidates;
search large historical corpora;
measure image similarity;
generate candidate mechanisms;
fit statistical models;
explore combinations faster than a human can.

Those capabilities are real. So is the danger.

A sufficiently flexible model can discover enormous numbers of correlations. Voynich is already correlation-rich. AI therefore makes the old Newbold problem more, not less, important.

The greater the pattern-finding power, the stronger the validation discipline must become.

Why AI Has Not Changed the Definition of Decipherment

An AI-generated translation is still a translation claim. It must still be reproducible. The mapping must remain stable. Grammar must generalise. Unseen passages must work. Illustrations should agree where the method predicts they should. Currier A/B must be explained. Exceptions must not require new rules whenever output becomes inconvenient.

Human or machine, the evidentiary burden remains the same.

See What a Real Voynich Decipherment Must Survive.

What Failed Solutions Gave the Field

It is too easy to write Voynich history as a comedy of errors.

That wastes the most valuable part of the history.

Failed solutions can produce new transcriptions, better scans, inventories of glyphs, comparative manuscript research, statistical datasets, falsification criteria, awareness of overfitting and entirely new historical questions.

Newbold failed as a decipherment but teaches the danger of overinterpreting microscopic variation.

Petersen did not solve the text but built infrastructure.

Friedman did not produce a final key but professionalised the cryptanalytic problem.

Currier did not translate one sentence but discovered internal heterogeneity.

Material scientists did not read Voynichese but eliminated entire historical fantasies.

This is progress.

The Question Has Become Harder—and Better

A century ago, a proposed reading could attract attention if a few words seemed plausible and the author identification was exciting.

Today a serious explanation inherits a much larger burden.

early-fifteenth-century material chronology;
codicology;
multiple visual regimes;
Currier A/B;
scribal variation;
line effects;
word families;
rare glyphs;
labels versus prose;
visual-text relationships;
local vocabulary;
frequency and entropy behaviour;
historically plausible cipher and abbreviation mechanisms;
reproducibility on unseen text.

This can feel as though the mystery has become worse.

It has actually become more constrained.

We know more ways a solution can fail. That is knowledge.

The Great Research Instruments of Voynich

Historical letters: establish early modern custody and reception.
Photography and photostats: allow the manuscript to travel intellectually without the object travelling physically.
Hand transcriptions: turn page impressions into comparable symbols.
Concordances: reveal repetition and frequency.
Cryptanalytic groups: compare mechanisms systematically.
Currier classification: reveals heterogeneity.
Digital transliteration: makes corpus-scale computation possible.
High-resolution digital images: expose small palaeographic and material details.
Radiocarbon dating: constrains material chronology.
Elemental and spectroscopic analysis: characterises inks and pigments.
Multispectral imaging: reveals information hidden from ordinary light.
Modern linguistics: separates language-like structure from language identification.
Information-theoretic and network methods: measure structure without requiring semantics first.
Digital palaeography: tests scribal and glyph variation across the codex.
AI: scales pattern comparison and mechanism exploration.

Notice what this list contains.

Almost none of the great advances is a claimed translation.

They are improvements in observation.

The Research History Is Also a History of Compression

The original manuscript is physically rich.

A photograph compresses it into visible light.

A transcription compresses it into glyph codes.

A word list compresses it into tokens.

A statistical model compresses it into distributions.

An embedding compresses it into coordinates.

Each representation makes some questions easier. Each loses something.

A token corpus can count q-series forms beautifully while forgetting that one token sat under a plant leaf and another ran around a zodiac ring. A visual model can compare page layouts while ignoring textual sequence. A materials analysis can identify pigment while knowing nothing about grammar.

Every Voynich instrument creates visibility and blindness at the same time.

That may be the most important lesson for the AI era.

Why the Next Breakthrough May Be a Better Dataset, Not a Translation

Imagine a future corpus that aligns every visible glyph with stroke uncertainty, line geometry, image boundaries, labels, bifolium relationships, scribal-hand probabilities, Currier state, material layer and diagram position.

No word has been translated yet.

But the manuscript has become a much better scientific object.

A real mechanism can now be tested across modalities rather than against one flattened text file.

Historically, this kind of infrastructure has repeatedly preceded major advances: Petersen before modern corpora; transliteration before modern linguistic modelling; digital images before digital palaeography.

The next breakthrough may again be a better question made possible by a better representation.

Primary School: Same Object, Better Tool

Give a child a tiny blurry photograph of a handwritten note and ask what they can see.

Then give them a high-resolution image.

The note did not change.

The available questions changed.

That is a century of Voynich research in miniature.

Secondary School: Build the Timeline by Instrument

rare-book access;
photography;
hand copying;
cryptanalytic tabulation;
digital transcription;
statistical computing;
radiocarbon dating;
spectroscopy and imaging;
digital palaeography;
machine learning and AI.

For each tool, ask one question:

What became measurable that was not measurable before?

This turns the history of knowledge into a history of changing observability rather than a list of famous names.

JC and Adult Readers: Separate Discovery From Interpretation

For every famous Voynich result, write two lines.

Discovery: what was measured or observed?

Interpretation: what explanation was proposed?

Example:

Discovery: two strong textual distributions exist.

Interpretation: they might be languages, dialects, scribal states, source differences or encoding regimes.

The distinction preserves good observations even when interpretations later change.

Reader Checklist: Before You Believe a New Voynich Breakthrough

What new evidence or instrument produced the claim?
Is the underlying data public?
Is the transcription assumption explicit?
Was the hypothesis formed before or after seeing the target examples?
Does the method reproduce earlier known constraints?
Does it explain Currier A/B?
Does it work on unseen text?
Does it preserve page geometry where relevant?
Does it distinguish observation from interpretation?
Can independent researchers reproduce it?
Does it explain failures?
Is AI producing a hypothesis, or validating one?

Frequently Asked Questions

Who first tried to solve the Voynich Manuscript?

The manuscript was already being investigated in the seventeenth century by owners such as Georg Baresch, who sought help from Athanasius Kircher. Modern mass scholarly attention followed Wilfrid Voynich’s acquisition in 1912.

What happened to Newbold’s solution?

Newbold’s elaborate Roger Bacon decipherment did not survive critical scrutiny. Its historical importance is partly methodological: it demonstrates how unstable microscopic or multi-stage interpretive rules can create convincing apparent readings.

Why is Theodore Petersen important?

He spent years making a detailed hand copy, annotations and concordance. That work helped transform the manuscript into a researchable dataset before digital transcription existed.

What did Currier discover?

He demonstrated strong differences between two textual populations now called Currier A and B and made important observations about scribal hands. This showed that the manuscript cannot safely be treated as one uniform corpus.

What is EVA?

The European Voynich Alphabet is a transliteration convention for representing visible Voynich glyphs with ordinary characters. Its letters are shape labels, not decoded phonetic values.

Has AI solved Voynich?

No accepted manuscript-wide decipherment has been established. AI can accelerate pattern analysis and hypothesis generation, but a translation still needs reproducible rules, unseen-text performance and coherent manuscript-wide explanation.

Research Foundations

The Final Idea

Voynich research has spent more than a century failing to produce one accepted translation.

That sentence is true.

It is also a terrible summary of the century.

The manuscript moved from rare-book curiosity to photographed object. From photographed object to hand-transcribed corpus. From corpus to statistical system. From statistical system back to physical codex. From one mysterious text to interacting textual, scribal, visual and material regimes. And now from a human-scale puzzle to a multimodal computational problem.

The answer remains unknown.

The problem is no longer vague.

A century of Voynich research did not merely fail to decode the manuscript. It taught us how much a real decoding would have to explain.

Voynich | Everything eduKate Knows and Tested | A Century of Voynich Research: How the Question Changed From Newbold to AI