A Possible Use for Generative AI
ThomasHairBrain > 11-07-2026, 01:59 AM
Hello,
This is my first post here and it is quite likely that this is either quite a bad idea or one that has already been tried (although I couldn't find it).
It also might be a bad idea as this thought was spurred by watching the most recent Voynich Talk video which calls out the use of AI here. I think this method may be distinct from those but if I get banned and the thread gets put in the slop drawer then I probably only have myself to blame.
The basic idea I have is what if someone trained a local AI model from scratch on the VM? A transcription of the manuscript could be fed into it for training and then see what new VM style text it can spit out. Of course we couldn't read this text any more than we could the VM but it might help to understand the characteristics of the text more, see what strings of Voynich script fed to it produce an output with a similar structure to the VM itself and what produce unstructured garbage. See if any sort of prompting can cause the script to break and see if it breaks in interesting and repeatable ways possibly across multiple models.
From what reading I've done it looks like the best approach would be using an LSTM model although even with that we have some possible issues of having material for the AI to work with.
Some possible concerns:
1. This might be a lot of time and effort chasing basically nothing and there might be little chance of this approach producing anything useful.
2. There is no universally agreed upon character inventory, there appears to be quite a high likelihood of positional variation playing a significant role in the Voynich script and an AI model could simply fail to take this into account. Alternatively variation in style could result in characters that are actually distinct being classified separately. Either one of these problems could result in gibberish or worse - real looking data that is misleading.
3. In order to make it work Latin characters (or other characters a computer can work with without a lot of adaption) would probably have to be used in substitution of Voynich characters, possibly causing issues as some human subjective thinking may be needed here. One approach would be to give every possible variant of a Voynich character a separate character but that could a) reduce the functional training data or b) produce more gibberish. Possibly could be mitigated with using several different competing models.
4. There might be insufficient data to produce anything meaningful even with a workable model.
So what do you lot think, worth pursuing and seeing if any interesting things can be discovered about the Voynich script this way or waste of time?