Protobufs don't follow this model. For example, a User message type will often be defined in one place and code generated from it in multiple languages. Every server and client uses essentially the same type, modified as suitable for that language. As a result, clients are usually working with types that they don't control and that have many more fields than they actually use.

Maybe clients should declare their own types containing just what they need and copy the data into them from the protobufs? But this does get tedious. On the other hand, in a schemaless JSON system, there is nothing statically checking that the clients and servers have compatible ideas of what a message should contain.