Out of something from science fiction.. would an "issue" like this ever be the type of problem where the AI can "understand" the cancer better than we can and see where a certain treatment might be more beneficial than we've thought of?
The example in the article given was that Watson recommended giving a blood thinning agent to someone with "severe bleeding", so I think, at least in this example, that there's a long way to go until Watson is ready for diagnosing without a doctor's consent and review.
Not to mention that even if the AI technology here was decent (which it doesn't appear to be) it suffered greatly from a dearth of samples. Machine learning requires big data, and cancer treatment data is, in many instances, just not big enough. An AI solution would need millions of well-labeled, perfectly formatted cases to learn from; there may be thousands of those for a particular subtype of cancer, but there's not millions.
Not if you understand mechanism. In reality the breakthrough will be biologists applying ML/AI, not the other way around. Domain and mechanistic knowledge are king.
Or is this just a plain old failure?