Are Human-generated Demonstrations Necessary for In-context Learning?