We created near-sentient algorithms — but now they’re devolving into bigots