Trust networks are needed: a graph of mutual endorsements. A trusts B, and B trusts C, so A can trust C. (Subject matter and degree dependant.)

The training data need to be acquired by "trusted" sources, which will vary per trust network.

Machines can't know truth without being able to make their own observations and so depends on the observations of humans and human adherence to truth. The same goes for goodness and beauty. They transcend humans, let alone machines.