Me too. Training a neural net of this scale is basically a lossy compression and promting just an interpolation of the data. But explain this to lawyers and courts...