0.8641
Herbal / pharmaceutical
Same scribal hand and Currier category. Different quires remain possible.
AI research / digital humanities / October 2026
I asked AI to read Voynich. The harder test changed the story.
The first result looked promising. Then I asked what else the algorithm might be recognising. A stricter comparison made the claim smaller and the research more useful.
Look closer.
The important finding
We found a pattern. We did not find a translation.
The Voynich manuscript is a medieval book whose writing has no accepted decipherment. We tested whether labels for its glyphs could predict which illustrated section a page belongs to. They sometimes could. When we also held the manuscript quire constant, the result fell to chance in a small sample. That changed what we could claim about the earlier scores.
01 / The question
What is the program actually learning?
My questions began with the original Yale scans, the drawings, and one line of text. I wanted to separate what we could see from what we might infer. An AI-assisted visual check of folio 2r produced matches and a specific glyph-label disagreement against two reference transliterations. It produced no meaning for the line.
We then moved to a broader test using the Ched-ai EVA transcription dataset. EVA letters label manuscript shapes. They are not translated words. A short Python program counted those labels, held entire folios out of training and tried to predict the illustrated section.
02 / Held-out folio evaluation
The result that changed the claim
0.9500
Herbal / biological
Same scribal hand and Currier category. Different quires remain possible.
0.5000
Herbal / pharmaceutical
Same hand, Currier category and quire. Two folios per class.
Scores are balanced accuracy, the average recall across two classes. A value of 0.50 is chance by this measure. The same-quire check contains only two folios per class, so it does not prove that subject matter has no relationship to the text. It does show why the higher scores cannot be treated as decoded meaning.
03 / The method
Make the correction visible.
A page's subject, position, layout and copying history can travel together. A classifier can recognise one of those signals and appear to recognise another. We held scribal hand and Currier category constant, then added quire as a stricter control. The second comparison changed our interpretation.
This builds on earlier research by Sterneck, Polish and Bowern, who found that text clusters track a combination of scribal hand and illustrated subject. Our contribution is a small reproducible example of why an exciting AI result needs a harder check.
What comes next
Curiosity, with a way to check it.
We have not identified a language, cipher key, plant species, author or translation. The next useful tests would compare independent transcriptions, hold layout and position more tightly, and ask any proposed meaning to predict unseen passages with outside evidence.
The published Space includes the method, the program and every recorded prediction. It does not redistribute the underlying transcription corpus.
Sources: Yale Beinecke Library, Ched-ai EVA dataset, Sterneck, Polish and Bowern.