Radiologist here with an interest in this topic. I think the problem with most AI applications in radiology thus far is that they simply don't add enough value to the system to gain widespread use. If something truly revolutionary comes along, and it causes a clinical benefit, healthcare systems will shift to adapt this in a few years. AI just hasn't lived up to it's promise, and I agree it's because most of the people involved don't get that the job of a radiologist is way more complex than they think it is.
Everytime I open a journal, I see more examples of either downright AI nonsense ('We used AI to detect COVID by the sounds of a cough') or stuff that's just cooked up in a lab somewhere for a publication ('Our algorithm can detect pathology X with an accuracy of 95%, here's our AUC').
Hyperbolic headlines - Geoff Hinton saying in 2016 that it's time to stop training radiologists springs to mind - don't help the over promise of AI, and then they shoot themselves in the foot when they underdeliver.
Earlier discussions about radiologists being self interested in sabotaging AI is tinfoil hat stuff - if I had an AI algorithm in the morning that could sort out the 20 lung nodules in a scan, or tell me which MS plaque is new in a field of 40, I'd be able to report twice as many scans and make twice as much money.
Companies come along every month promising their AI pixie dust is going to improve your life. It probably will, but 10 years from now, not today. The AI Rad companies are caught in an endless hype cycle of overpromising and under delivering.