You brought up the example of radiologist. My take-Radiology is IMO a great candidate for AI, even the Markov chain flavored version we have now, because it's primarily pattern recognition. AIs today recognize patterns.

I still say that even if you limit the scope of AI use to pattern recognition type tasks, that is still a huge chunk of human professional expertise.

Accountants recognize patterns in revenue and spending. Doctors diagnose based on observation of patterns. Law enforcement recognizes patterns of behavior to determine law application. I really don't think that much of what we consider professional tasks is truly creative or demands synthesis.

Training an AI on the output of an AI will always fail. Too much of the training set nowadays is polluted with AI output.

Which magnifies the essential problem - AI gets trained on a steadily declining base of original knowledge and practice. When the trainers stop training, then the patterns known to AI will deteriorate as systems are replaced.

Entropy will destroy most human expertise. All we will have left will be recordings and memories.