As someone who did a PhD and published and read many papers (in ML) I believe vast majority of papers in ML are misleading. If in fact every paper that claims a better performance over state of the art was true we would have solved AI by now. You see all sorts of problems when you dig deeper into the technical details of peer reviewed publications (even in top tier conferences) including misleading baselines, statist…
How do you measure consensus? There are many researchers who claim there's a consensus of whatever they happen to believe, but when counter-examples are pointed out they start No True Scotsmanning ("no real expert believes..."). Health related research is in a much worse state than ML unfortunately. ML suffers from metrics gaming and overfitting but there's probably not much outright fraud? Common estimates are that…
Odd that tenure is singled out as the core of the problem. I'd say the undermining of tenure is a large part of the problem.
Research is hard - very hard - you are being asked to discover stuff nobody else has - to push the boundaries of knowledge. I would argue that the idea that one failed research project, and you are out on your ear because of a gap in publishing - is a core part of the problem.
In terms of medical doctors doing research - I'd argue that's a special dangerous mix - medical doctors need to have an element of self-belief just to be able to do their job ( life and death decisions ). Instilling it is part of the training.
That self-belief, plus a lack ( in most cases ) of any in depth scientific research training is a dangerous mix.
Medical training != science training.
In this specific case I have no idea about the rights and wrongs.