This was my opinion circa 2011 or so when I was recapping the old AI.

Consider, for instance, the successful early medical diagnosis program MYCIN

https://en.wikipedia.org/wiki/Mycin

which like any kind of diagnosis process is a problem of reasoning with probability. Language understanding has the same issue, like if you wrote a grammar for English you'd find that common sentences have 1000s of possible ways to parse and you will need to either make a guess or keep your options open.

MYCIN had a half-baked approach to reasoning about uncertainty that worked, one of the reasons why symbolic AI fell out of favor was that nobody developed a generally useful approach to bolt probabilities onto logic.