Legally, https://towardsdatascience.com/the-most-important-supreme-co... gives a good example how transformational machine learning classifiers generally fall under fair use. It does raise a good point that generative machine learning, like this, has not been explored legally yet.

This is still research which will likely provide public good if/when they publish results and methods. Probably, they'll do a different dataset for any commercial work given the profanity problem highlighted in the article.