Could the Voynich text be made up? People who write nonsense by hand don't write random letters. Gaskell and Bowern (2022) asked 42 volunteers to write made-up text and found it can look a lot like a real language. So the fair test isn't the Voynich against random text. It's the Voynich against text people made up. We ran each test on five kinds of text.
1. Does it look like made-up text, letter by letter and word by word?
Gaskell and Bowern's 42 measures (repeats, where letters and words fall in a line, how well it compresses, and more), with a classifier trained on their real texts and their made-up samples. Each dot is one text. Further right means more like made-up text.
2. Long-range order
Whether long and short words cluster (word-length autocorrelation over the whole text), and how often a pair of words comes back.
3. Words that keep to their topic
In a real book, some words gather on the pages about one thing. Montemurro and Zanette's measure counts the words that keep to fewer pages than chance would put them on. The made-up samples are too short for this test (Gaskell and Bowern's own limit), so it runs on everything else, cut into pages the length of the Voynich's.
4. Words tied to pictures
If a word means something drawn, the pages it appears on should share drawings more than chance. Here, pages' drawings are shuffled within their section 200 times.
What this can and can't show
No test can prove a text has no meaning, only fail to find it. And a text can be meaningful and still score like made-up text if its writing system hides its language. The made-up samples are short, and were written by volunteers, most of them Yale students in 2018 and 2019, not by a 1400s scribe. Every number here comes from the development pages; the pages kept for a final test stay closed.