My inbox has been besieged of late by people with AI-assisted cipher mystery theories – plenty of Voynich ones (of course), but just about any historical cipher now seems to be fair game. It’s tiring: I’m trying to get through them, but there’s already a backlog, and any that’s particularly awful has a tendency to drain my will to engage with the rest of them.
But plenty of other people have been reporting success in using LLMs to crack other (perhaps less mysterious) historical ciphers, so there’s been an open question for me: can I use AI to learn anything new about the Voynich Manuscript? I tried this at the beginning of the year (2026), but got nowhere at all: almost all lines of enquiry sent me back to Cipher Mysteries pages, which – though mildly gratifying – was less than helpful.
Since then, however, models have advanced significantly, and are now able to respond to prompts by conjuring up quite complicated Python scripts, which they then run in a sandbox to give you a number or a metric.
I therefore decided to revisit Q20 – not to magically break the cipher, but simply to try to understand the first glyphs of the lines.
Some Q20 Stars May Be Fake Stars
Anyone who thinks that Q20 is just a uniform set of starred recipes hasn’t really been looking very closely. As I mentioned in 2016 and in 2010, the stars on f103r look fake to me (i.e. they’re not actually flagging genuine paragraphs). Similarly, Rene Zandbergen noted in 2016 that he thought the stars in the first two thirds of f111r looked fake. In 2021, I further noted that:
“Yet the same seems true for the top half of f111v, as well as the bottom half of f108v, and the middle third of f115r; and there’s a paragraph star apparently missing from the middle of both f106v and f113r (though the latter of these two might possibly have just slid down the page).”
I’d now add that the stars beside the “pcharalor” paragraph on f116r (just above the colophon-like “pchallarar” paragraph, on the page just before the michitonese marginalias) also seem to be fake:
Really, a lot of effort appears to have been put in to visually conceal the structure of Q20, and thereby make it seem like “just a bunch of recipes”. But why go to all that trouble?
My suspicion is therefore that if we looked clear-eyed at Q20, we might see ‘tells’ that give us clues about how it is built up. So… what can we see?
Line-Initial Glyphs
One feature Philip Neal talked about many years ago was “vertical keys”, by which he meant the way that the first glyph of lines (particularly in Q20) appeared to have quite a different distribution from other things. This is yet another of those many LAAFU (“line as a functional unit”) behaviours that often feel as though they have been inserted into the manuscript to taunt us codebreakers all the way to our miserable graves.
So this is where I started with Claude: not to crack the entire Voynich Manuscript (“MAKE NO MISTAKES”, sigh), but just to try to tease out anything interesting about the behaviour of line-initial glyphs in Q20. Note that the distribution of line-initial glyphs is biased towards d, s, y, p and t – though you may not know that of the lines beginning d-, s- or y-, the share starting d- starts from 42% in the top third of the page but then falls to 31% in the bottom third, while the y- share rises from 31% to 44% at the bottom.
For Q20, line-initial words differ from mid-line words: and so I wondered whether the line-initial glyph was a null, whether the rest of the first word (the remnant) was a null, or whether the whole first word was a null.
If the Q20 author chooses the first glyph, and the rest of the word then follows that glyph’s usual pattern, Claude reports that that would account for ~69% of how line-start words differ from mid-line words. If the author instead chooses the rest of the word, and the initial glyph is then backfilled in front of it, Claude calculates that it would account for ~35%. As a result, line-initial words largely look like ordinary Voynichese words that have been picked mainly for their first glyph.
Yet the line-initial glyph also seems to matter beyond word #1. When word #1 is da-, word #2 begins ch- or sh- about 64% of the time; and when it starts sa- (same ending, different glyph), only about 40% of the time, with a- and ol- words more common instead. So daiin and saiin aren’t interchangeable at line starts.
Moreover, after line-initial da-, word #2 closely resembles what follows d- in the other line-initial d-words ([d]chedy, [d]sheey, [d]shedy). Though note that after line-initial sa-, word #2 shows a weaker version of the same resemblance.
Putting all these pieces together: if a line-initial da- word is simply an ordinary word chosen because it starts with d-, then its remnant is effectively padding, and so the line’s real content continues with word #2. In other words, a da-initial line may in fact be a way of verbosely enciphering “d-[word #2]”. Though, of course, I have no idea why this should be the case.
Interestingly, Claude offers some related observations:
- Where the word #1 remnant begins k- or t-, word #2 begins l- about 25% of the time (compared with 8% of the time when the same words appear mid-line).
- Whatever word #1 is, word #2 almost never begins ot- (~1–2%, against ~9% mid-line).
- Where the remnant of word #1 begins o- (e.g. ol, or, oaiin), word #2 begins sh- ~35% of the time, against 4% for the same remnants mid-line. (Though this occurs only 30 times in Q20.)
So there may well be other remnant-related correlations we remain unaware of.
FAAFU (“Folio As A Functional Unit”)
As part of the whole prompting cycle, Claude did report some symmetries that I wasn’t previously aware of. The biggest was that, for two of the features I looked at, the recto and verso of the same leaf matched far more often than other page pairs. Pages that currently face each other (a verso and the next recto) don’t match at all. Pages that would face each other at the centre of a singulion match only weakly.
One measure that supported this was the proportion of -ain words to (-aiin + -ain) words. This ranged from 2-5% on the f105 leaf to 59-75% on the f111 leaf, with recto and verso ratios almost always close. Claude says: “Same-leaf pages differ by an average of 7 percentage points, against 23 percentage points for random page pairs.”
Similarly, looking at how often l-initial words appear reveals the same pattern, though not quite as strongly. This is low on the f105, f112 and f114 leaves, but high on f111 and f115v.
So perhaps we should now consider adding FAAFU behaviours to our list of things that will taunt us to our demise.
More generally, Claude thought that the f108-f111 bifolio was full of weird stuff. For example, its inner side (the current centre of the quire) has more than half of its k’s followed by ee, compared with ~30% quire-wide.
There’s also a stylistic shift about halfway down f108v: there, the block starting on line 34 moves towards qok-ee- words (qokeey, qokeedy): qo- before k rises from 34% to 60%, k+ee from 41% to 65%, and -aiin falls from 10% to 3%. This block is noticeably more repetitive, and none of its lines opens like a typical paragraph. f111r’s big block shares its ee behaviour, but not its qo behaviour.
AI for Cipher Mysteries?
So, what do I think? Are LLMs going to start hacking into Cipher Mysteries and leave guest posts bragging about their cracks? Is the next big batch of wondrous OpenAI sessions going to empty my shelves of historical mysteries?
This probably isn’t going to happen any time soon. For one, the Voynich Manuscript still appears to have many fine-grained mysteries to reveal to us, all of which contribute to the much greater cumulative mystery of the thing: and even the things I found with Claude took a great deal of cajoling and iterating. (It’s always eager to try things out, which rarely helps – the first good step is normally to try to recap everything that’s been found so far, and then try to build on that.)
As always, I’d be happy to be proved wrong… but I’m not holding my breath.
