I'm not sure whether to take this is a definitional statement or as a conclusion based on observation.

If it's definitional in that structurally-what-generative-AI-does-is-necessarily-syntax, how do we actually delineate syntax from semantics? Would we say that the problem with advocating for two contradictory positions is syntactical in nature? That seems...odd. And yet, if you consider sufficiently complex sets of relationships, maybe that's true: it can be cast as a syntactic problem rather than a semantic one. But then doesn't that beg the question of whether semantics is actually different in kind, or whether it's really just latent in syntax (to use ML lingo)?

If you are instead drawing a conclusion based on observation, are you familiar with the Anthropic Golden Gate Bridge model (which, when they had it up, would (almost) always find some way to work the Golden Gate Bridge into its answers, despite the answers otherwise seeming pretty decent), or the more general result about abstract features (https://www.anthropic.com/research/mapping-mind-language-model)? If we were to find that, or to be able to make manipulations with those consequences, in humans, I surely wouldn't interpret this as "evidence of a neural correlate of syntax". I would call that semantics. Wouldn't you?

Finally, I don't think we can conclude that general intelligence requires consciousness, so I don't know that I can agree with "...consciousness (and therefore understanding)...". In us, yes, understanding requires consciousness (as we define the terms). But to what extent that is an artifact of how we couple our capacity to select high-efficacy models ("understanding") to our running evaluation-goal-determination-attention-action-selection process ("consciousness") is not clear to me. Even that seems slightly dubious to me; if I suddenly have a realization and "get it", I am of course aware that I "got it", but it seems like the getting it might be prior and the conscious bit is only the awareness of the understanding, not the understanding.