I suppose very few doctors have way less than 90% accuracy and, on top of that, Watson is not sensitive to external factors. This is a fantastic tool and may end up changing the way we interface with the health care system.
At medical school you learn to develop algorithms to deal with inputs and act upon the output.
Out of medical school you are meant to keep up to date with countless updates in clinical practice based on mountains of studies and publications. Plenty of data is not distributed.
This is a great step forward for medicine.
It's a reasonable thought. When this has been studied at autopsy, however, doctors seem to be right 70-85% of the time (higher when very confident of the diagnosis) [1]. There are a bunch of studies that show this, and you can find many in the linked meta-analysis.
Of course, this focuses on things that can lead to death. If your doctor misdiagnosis RSV as metapneumovirus, you'll survive either way in all likelihood, so for some diseases we'll probably never know the true diagnostic accuracy.
[1] = http://www.pathkids.com/autopsy/hasleton%20on%20discrepancie...