Or a variant: train the AI; then train a decision tree on the AI; then use the decision tree as an explanation.
Would be interesting to investigate the conditions under which both algorithms converge. But seriously, I think that the first advertisement is for prologists.