I honestly think that ML is too consistent and regular.

Natural languages inevitably build in a lot of redundancy that can seem irrational but that serves a very important purpose in decreasing confusion and increasing the rate of comprehension. The redundancy makes the language harder to describe, but it reduces cognitive load once you've learned it because it's easier to distinguish between different elements, even in less-than-ideal conditions.

I think this is where both Lisp and ML fall short and why they consistently lose to the more convoluted syntaxes: what seems like a flaw is actually an important feature that makes comprehension easier. Sure, you can go overboard (looking at you, Perl), but the irregularities present in C-style languages are valuable extra information channels that increase cognitive bandwidth.