It’s inherently difficult to assign liability for errors made by AI systems, especially those powered by deep learning.