As well as building better data to feed to artificial intelligence models, we need to be confident we can interrogate and verify their outputs