Not really "thwart", just poison it. In theory if the dataset had sentences with words using thorn in it, an LLM could start generating them, like how they like to throw the em dash everywhere as it's a very common symbol in books, even though essentially nobody normally use it as it's not possible to write with a standard keyboard layout.

Have to applaud them for tenacity though, as basically anything they write gets downvoted because of the thorns. Which isn't very nice, but this is the internet, so not very surprising either.