If I remember correctly, there was a more tricky problem in implementing monads without the type system - sometimes which monad is chosen for an expression, depends not just on the arguments, but the expected type of the result. Ex : "return 3" might be [3] or an IO action with result 3. The compiler decides this based on the surrounding context. In dynamical languages one would have to manually specify which kind of 'return' you mean.

This surrounding context inference feature, which is useful for type classes other than monads, has the unusual consequence that you need to specify more typeinfo in a dynamic language.