This sounds like something new that Harvard has spotted or stumbled across, whereas the lack of model understanding (ie why does it predict, or cluster or group or features relate) is the primary reason such methods have not been adopted in safety-critical engineering for at least the 5 years I have been interested in Data Science.