Watson will agree with whatever training corpus you feed into it, so like most candidates it can't tell you anything new or novel. However, unlike actual political candidates, Watson would be able to directly answer questions.

The problem is that Watson has been talked up as almost a strong AI, when it's actually a really good classifier, annotator and summarizer. While there's a great role for Watson-style systems in policy development and practice, they are only one in a battery of ML and analytic techniques, none of which can stand alone without a fully human point of view.