Not as a rule though - so many ML systems are utilizing data that is streaming in from passive sensors or transactional streams and is not human curated at all. The human aspect isn’t an intrinsic property of ML or even these algorithms only particular applications (and I would guess a minority of applications too).

Given that, it seems to be a clear miss to apply that logic generally. I have to believe it most likely stems from a lack of basic understanding and competence on the authors part.

Edit: Here are two examples that I have personally worked on:

Global calculations of weather composite reflectivity using 20 years of historical satellite imaging data from NASA

Superior RF demodulation (under certain circumstances) using RF transmissions as received and sent.

Both of these utilize modern ML imaging models, neither require any human labeling only streaming data (which in these cases began collection long before modern ML techniques were in widespread use). The applications in the natural sciences are endless not to mention the applications more on the business intelligence side using transactional data. Only in specific cases is human labeling required but because of the high cost of that data it is by its nature dwarfed by that which is collected naturally (not to mention often error prone). It is for that reason that techniques to ingest data that is more and more natural to collect are growing in favor.