AI Finally Translated The Voynich Manuscript — It Isn't A Language At All

Scientific concepts, discoveries, and nature/phenomena discussed

Origin, physical dating, and artifact verification

• Radiocarbon dating of vellum (University of Arizona, 2009) established the manuscript’s writing material dates to roughly 1404–1438, supporting that it is an early-1400s artifact rather than a modern forgery.
• The manuscript contains:
  • ~240 surviving pages (believed originally closer to ~272)
  • Invented script (no matching known alphabet)
  • Unidentified plant illustrations
  • Unfamiliar star-chart constellations/sky drawings
  • Human figures with unclear interpretation

Handwriting / textual production hypotheses

• Paleography (researchers who study handwriting) suggests the text may have been produced by multiple scribes using a shared invented alphabet, inferred from consistent letter-shape habits across pages.

---

Computational/AI language identification claims and critiques (2014–2017)

1) “Plant/constellation word readings” via a historical-pattern matching method (2014)

• Stephen Bax (2014) proposed partial readings by borrowing a method used in ancient Egyptian hieroglyph decipherment:
  • Start from proper-name-like elements
  • Expand outward to other symbols using associations with nearby illustrations
• Reported tentative matches (as stated in subtitles):
  • Plant names including juniper and coriander
  • Possible constellation association: Taurus
• Criticism:
  • Other researchers argued the method was open-ended and could fit many language families (not a strong language-unique decipherment).

2) AI “Hebrew” detection using alphagram/anagram-style scrambling (2017)

• Core idea: Voynich words might be alphagrams—letters rearranged into alphabetical order (with vowels removed in the proposed interpretation).
• Training/validation step (NLP test unrelated to Voynich at first):
  • Train on the Universal Declaration of Human Rights translated into 380 languages
  • Shuffle letters within each word into alphabetical order
  • Test whether the model can still identify the source language
  • Reported performance: ~97% correctly identifying language
• Application to Voynich text:
  • The system ranked Hebrew highest among candidates.
  • Multiple languages were considered, but Hebrew was flagged most likely.
• Major critique:
  • The translated output required manual spelling corrections and Google Translate to become coherent.
  • Medieval-language experts argued an algorithm trained on modern languages cannot reliably infer the language of a 1400s technical text; reliance on translation tools reduces credibility.
• Researchers emphasized they did not have a full, proven decipherment.

---

Statistical tests for “does it behave like language?”

These approaches do not claim a specific translation, but test whether the text’s statistical structure resembles natural language rather than random noise or artificial word generation.

1) Hidden Markov model to map two writing “styles” (2011)

• Kevin Knight & Shravana Reddy (2011) used a hidden Markov model (HMM):
  • Goal: classify each page as leaning toward Currier “Language A” vs “Language B”
  • Result: the model generated a page-by-page style-strength map
  • Correlation: Language A dominated herbal sections; Language B dominated biological/recipe sections.
• Additional key observation:
  • Low entropy in letter patterns (predictable letter ordering relative to real languages like English/Latin), dividing researchers into:
    • those who view it as evidence of an unusual but real language
    • those who view it as a warning of artificial text generation

2) Information-theoretic “network” clustering of content words (2013)

• Marcelo Montemurro et al. (PLOS ONE, June 2013) treated the manuscript as a network of connected words.
• Information theory claim:
  • In natural languages, content words (nouns/verbs) show strong clustering around topics and act like network “hubs,” while small connector words distribute more evenly.
• Result:
  • Voynich word usage patterns were reported as compatible with meaningful language (not random gibberish), without proving specific semantics.

3) Multi-test statistical physics comparison to real languages (2013)

• Diego Amancio et al. (PLOS ONE, 2013) compared Voynich against a large corpus of confirmed human languages using three tests:
  • Zipf’s law (frequency distribution: few words very common, most rare)
  • Complex network shape of word relationships
  • Time-series-like repetition tracking across page order (pattern of word recurrence/clustering)
• Result:
  • The manuscript’s statistics fell within ranges produced by real languages, not random noise.
• Limitation:
  • The framework could not identify which language or prove hidden meaning.

4) Large-scale entropy/character predictability comparison using Wikipedia corpora (2020)

• Luke Lindemann & Claire Bowern (2020) scaled comparisons using:
  • Cleaned Wikipedia text across many modern languages
  • A separate historical multilingual collection spanning centuries
• They measured conditional character entropy:
  • predictability of a letter given the preceding letter(s)
• Result:
  • Voynich characters had unusually low entropy (more predictable than normal English/Latin-like patterns in tested corpora).
• Important nuance:
  • Certain real encoding methods—especially those using bigraphs (two letters combined into one symbol)—can also drive entropy down into a similar low range.
  • Interpretation offered: if real language exists beneath, it may have been processed through unusual encoding.

---

Physical “earlier attempt” evidence (2024)

• Lisa Fagan Davis (2024) analyzed multispectral images (from the earlier Lazarus Project).
• Discovery:
  • Faint margin notes containing a string of Roman letters.
  • Interpreted as likely left by someone attempting an early decipherment.
• Significance:
  • Confirms the manuscript has been treated seriously as a solvable puzzle long before modern computational methods.

---

Competing “not a real language” generation theories (mechanical/algorithmic hoax models)

1) Cardan grille + syllable tables generator (Cryptologia; researcher Gordon Rugg)

• Gordon Rugg (Keele University) proposed a 15th-century-capable low-tech mechanical generation:
  • Cardan grille (masking/reveal pattern device)
  • Pre-made tables of word-building syllables
• Claim:
  • One could generate endless plausible-looking invented words without encoding meaningful hidden language.
• Rugg (2016 follow-up):
  • argued the method could reproduce several statistical “fingerprints” identified by earlier computational studies.

2) Self-referencing mechanical word construction (Cryptologia; Torsten Timm & Andreas Schinner)

• Torsten Timm & Andreas Schinner proposed a mechanism where:
  • new words are built from pieces of previously written words
  • identified from how word spellings repeat/shift across the manuscript
• Claim:
  • Statistical patterns could arise as byproducts of mechanical generation, not real semantics.

---

Examples of overhyped/unverified “AI solved it” claims (warning about methodology)

• Gerard Cheshire (University of Bristol, 2019): claimed “proto-romance” decipherment; quickly rejected and a press release was retracted within a day.
• Other viral claims (not peer-reviewed):
  • “old Turkish”
  • astrology/medical treatment interpretations attributed to a specific monk
• The subtitles emphasize these claims lacked peer-reviewed publication and independent reproducibility.

---

Lists / methodology outlines (as presented)

Kondrak & Hauer (2017) alphagram-style NLP pipeline

• Train an NLP classifier on Universal Declaration of Human Rights translations in 380 languages
• Create scrambled training/evaluation inputs by:
  • shuffling letters within each word into alphabetical order
• Validate:
  • evaluate ability to identify source language despite within-word scrambling (~97% reported)
• Apply the trained system to the Voynich manuscript text
• Interpret the top-ranked language as the likely source under the alphagram hypothesis (Hebrew reported)

Knight & Reddy (2011) hidden Markov model classification

• Use a hidden Markov model over a transcribed version of the manuscript
• Goal:
  • assign each page a probability/strength of belonging to:
    • Currier Language A
    • Currier Language B
• Output:
  • a page-by-page statistical map matching herbal vs. other sections
• Additional analysis:
  • observe low entropy in letter/pattern behavior

Montemurro et al. (2013) information-theoretic word network test

• Represent manuscript as a word network (connected words)
• Use tools from information theory
• Compare network behavior to expectations from real languages:
  • meaningful content words cluster/hub around topics
  • connector words distribute broadly
• Conclusion framed as:
  • patterns compatible with meaningful language, not random noise

Amancio et al. (2013) statistical physics comparison set

• Compare Voynich vs. large corpora of known human languages using:
  • Zipf’s law
  • complex network structure
  • sequential/repetition “time series” behavior
• Conclude:
  • Voynich lies within statistical ranges typical of real languages, not random noise
• Limitations emphasized:
  • cannot identify language or prove hidden semantic content

Lindemann & Bowern (2020) conditional character entropy scaling

• Build multilingual corpora from:
  • Wikipedia (modern languages)
  • historical text collections
• Compute conditional character entropy:
  • how predictable a letter is given preceding context
• Compare Voynich distribution to corpora
• Note encoding hypothesis:
  • encoding methods like bigraphs can yield similarly low entropy

---

Researchers or sources featured (named in the subtitles)

• Wilfrid Voynich
• Johannes Marcus Marci
• Rudolf II (Emperor)
• Athanasius Kircher
• Georg Baresch
• Stephen Bax (University of Bedfordshire)
• Gordon Rugg (Keele University)
• Grzegorz Kondrak (University of Alberta)
• Bradley Hauer (University of Alberta)
• Lisa Fagin Davis (executive director, Medieval Academy of America)
• Damien Fleming
• Kevin Knight (computational linguist)
• Shravana Reddy
• Prescott Currier (Currier “Language A/B” styles)
• Marcelo Montemurro (PLOS ONE team)
• Diego Amancio (University of São Paulo; PLOS ONE team)
• Luke Lindemann (Yale)
• Claire Bowern (Yale)
• Torsten Timm
• Andreas Schinner
• Gerard Cheshire (University of Bristol)
• Lazarus Project (source of multispectral images)
• The Times of Israel (media source quoted)
• Google Translate (tool used in critique context)
• Universal Declaration of Human Rights (source text used for training in the NLP test)

---

Video topic takeaway (“scientific standing” as described)

• The manuscript is treated as a real early-1400s artifact, with multiple computational studies showing statistical compatibility with natural language behavior, while multiple mechanical-generation and decoding critiques argue those statistical signals could be produced without real underlying meaning.
• No approach described in the subtitles yields a confirmed translated sentence accepted by the wider expert community.