The name "garden path" reminds me that during my Turbo Pascal loving 80s I played with making “fog” from markov chains. The idea was to computer-generate reasonable-reading nonsense text and I'm sure you've all tried it as kids too! :)

Of course my naive approach suffered from lack of corpus; if it picked a rare word, it’d get railroaded into using the very limited set of words that follow that rare word and so on.

I remember running it - this was before the internet days, so I can’t imagine where I got the input text from - and getting back a page-full of gobbledygook. And there right in the middle, somehow drawing the eye like a magnet, was one of those railroaded sentences:

herbaceous

Well when your random roulette picks the word herbaceous, its likely only seen it once or so in its learning text, so there aren’t so many words it can pick next:

herbaceous border

So far so good. But where does this lead next?

herbaceous border disputes

Love it!