For example, if you think mankind might someday create an AI that would satisfy the preferences of humans, then you need these preferences to be expressible mathematically, as some kind of utility function. If you're a non-realist, you're okay with that, as long as the utility function is specific to humans (i.e. the utility functions of human-built AI and alien-built AI don't have to be the same). But if you're also a non-cognitivist, then you're in a difficult spot, because obviously one can make true and false statements about a mathematical function. Does that make sense?

(Disclaimer - I'm a Google engineer as well, and also pretty active on lesswrong.com which has covered these issues in much detail.)