[thaumasiotes]

> The way it was explained to me by a consultant was to qualify there had to be technological risk of failure (i.e. it is unknown whether or not it is possible to do task X via software).

Ok...

> However the first few companies using deep learning to do facial recognition from photos (eg. facebook automatic tagging)... well they would have faced technological risk. It wasn't known at the outset that deep learning would produce sufficient quality matches to be useful.

This is a wildly different standard. It certainly was known at the outset that facial recognition from photographs is possible to do in software, because people do it all the time. They just didn't know how.