Maybe we reckon with the issue that "understand" doesn't mean anything in the absence of it's implications and we rephrase the question. What do you view as consequences of deciding that it "understands"?

This might go with the question of whether the Turing Test is a valid measure of "intelligence". If we can't tell the difference between a machine understanding something and a person, is that all that is required?

There's a view that intelligence without context is meaningless. (You can't do pattern matching without definition of what your signal is -- and if you just use a trivial measure, like entropy, then white noise is the best signal!) So, you can't be generally intelligent. You can be great at a lot of things, but that means you have priors for those things, which inherently predisposes you against other priors. (Though there are of course basically infinite dimensions, in which you can have priors, but let's just assume some dimensions are more important than others.)

What we want is adaptability under goal permanence, meta-goals, usually called alignment. We want an intelligence that is very similar to us. And we like to talk a lot, and we can talk about everything, so we can represent our intelligence through talking, so a Turing test could be a great way to measure how human-similar two "intelligences" are.

The machine has no reference for its terms, "salt" does not refer.

To a machine "salt" is just a distributional pattern within a large body of text, it isnt salt.

Ie., what computer scientists call "semantics" is again metaphorical BS, as much as calling an if-statement a "decision maker".

Denotation is the most fundamental part of what a natural language is (ie., it is about the world, and caused by the world in a fundamental way). Machines processing terms without their denotations are not using language, they are shuffling it aorund.