The Hindley-Milner "infer as much as possible" approach presents jagged-seeming edges to the user experience because the "possible" part -- restriction imposed not by language authors, but by the logic itself -- isn't always intuitive.
Artificial limitations can provide comfortable assurance. The no-inference of Java or minimal-inference of Scala give you less, but the limitations are more predictable. You don't even try giving the inferencer anything complicated, so you don't create opportunities for surprise.
In Haskell we have so much inferencing power it's easy for us to push the inferencer really far without even thinking about it. I suspect we ought to be in the habit of adding more type annotations than most of us do.
