Imagine a language where you work with blobs of data and have no restrictions on the operations between them. Like assembly languages. You can add a char* and float and the compiler won't complain. I would call that untyped. A typed language restricts the operations you can perform on blobs of data by classifying them.

A user could build a more restrictive type system on top of an already typed language. The distinction between a type system and an arbitrary source of runtime errors is that a type system has the properties i described earlier, and also it has the cohesion you would expect from it being a "system". A smattering of if statements wouldn't qualify.

That's one way to look at it I guess. Everyone has their own working definitions so I still it's still reasonable if you consider dynamically typed languages to be untyped.