Machine learning etc could be great for medicine but the profit motive means that the best bullshitter will get these early contracts. The best bullshitter has an actual product and it looks great in demos but they cut corners so its fuckups are difficult to audit and were swept under the rug when the few engineers that understood the problem raised the issue.

With very carefully collected and annotated datasets and heuristic guard rails this stuff could actually be great. But those things are slow and expensive.