There are things like too-clever type systems that get into your way, instead of catching errors. Constness can be silly in C++ if you are not careful, but that's nothing against the situation in e.g. Haskell.
Then there is type inference, which really scares me. When I write (in pseudocode):
fun add(a, b) { return a + b; }
the meaning can change depending on how this is called at a different place in my code (e.g. with strings or ints as arguments). On the one hand, this confuses me - if there is a conflict, it can be incredibly hard to debug. On the other hand, the compiler can get confused. If I call add(get_number(x), get_number(y)), I know that a and b are going to be ints, but the compiler might be unable to infer this, depending on how complex get_number is. Type inference is magic, and IMHO often not worth the saved keystrokes. I'd rather have my language use a Variant type when no type is specified (and sure, when it can infer the type, use it as an optimization).