Point 1 draws parallels with machine learning's pixel attacks. That one pixel doesn't define the image, but it happens that changing it changes the classification of the whole