At best, you may be able to justify black boxes providing secondary indicators: Maybe using AI to study cancer detection might lead you to a new solid discovery, but "we use AI to determine if you have cancer" should never be the mission, as it fails to generate useful information about how it is detected.