If this were truly a blackhat effort, it could improve the quality of empirical papers at conferences specifically by practically forcing a reproducible-by-the-conference-XYZ org of the results on some benchmark using a Docker image, etc etc.
On the other hand, for investigative or more purely academic research, this could be another layer of nightmarish crap that people would have to sift through.
Also it looks like at least one example is basically cloned from the source set, just with...basically greatly reduced recall of the content (beyond some high points).
This feels nightmarish, to be honest. Well, I guess we'll find one way to adapt to it, or simply fall back to certain, much more provably-analog methods of doing things....
Real: https://arxiv.org/pdf/1409.6041.pdf
Fake: http://arxivgen.com/pdfs/domain-adaptive_neur-3p5.pdf
