> The map of London (where I live) that I use to navigate on my phone is not in any sense the best map of London that has ever been created. It is definitely not an improvement on existing understanding, but it's a great improvement on not having a map.

But if the "map" is actually composed by a ML process parsing texts about London and arranging the place names based on proximity of mentions and cardinal direction references (this might be a fun exercise, might actually result in something that looks surprisingly like a map with the right data sources, and sufficient tuning, and absolutely won't have any real world use), people are going to ask why you trusted it.

To be fair to the authors of the original paper, their actual claims appear to fall some way short of the blog's insinuation that they were proposing that academics should actually use their model decide to spend hundreds of hours of work replicating research papers (which would be the perfect example of misuse of ML). But what they've actually done is a very basic statistical ranking of possibly misclassified subfields without any ability to address systematic error.