In at least some cases, it would be considered fair use. The wikipedia page on fair use has this to say, regarding Google's book scanning for example: As part

In at least some cases, it would be considered fair use. The wikipedia page on fair use has this to say, regarding Google's book scanning for example: As part of the ruling that found the book digitisation project was fair use, the judge stated "Google Books is also transformative in the sense that it has transformed book text into data for purposes of substantive research, including data mining and text mining in new areas" [1] To me it feels like a continuum. On one end of the scale, the model has been informed by copyrighted data. The model has learned something about the structure of the data, which generalizes to new examples easily. On the other end of the spectrum it's just an efficient encoding of the corpus dataset (in this case mappings of sounds to text). It's not learned anything... How do you distinguish these cases? Can the second type of "machine learning" be used to circumvent copyright in some cases, and is it? [1] https://en.wikipedia.org/wiki/Fair_use#Fair_use_in_particular_areas https://en.wikipedia.org/wiki/Fair_use#Fair_use_in_particula...