The article doesn't justify the title. Neither the oncologists nor the model were asked to predict when death would occur nor were they asked to predict if death would occur within the next 90 days.
Instead the phrasing is very odd leaving it open to interpretation for the humans as being "surprised" or "not surprised". Why not simply predict death or no death?
I suspect it's the difference between soliciting a 'most likely' versus 'within some upper margin of confidence/error' type answer from the humans. The model appears to produce a different output entirely that has to be adapted to meet what the humans were asked.
For a fair comparison, need both parties need to make the same type of prediction.