> There was for years an experimental period, when ethical and legal considerations about where and how to acquire training data for hungry experimental models were treated as afterthoughts.

Those things were afterthoughts because for the most part the experimental methods sucked compared to the real thing. If we were in mid 2016 and your LSTM was barely stringing together coherent sentences, it was a curiosity but not a serious competitor to StackOverflow.

I say this not because I don’t think law/ethics are important in the abstract, but because they only became relevant after significant technological improvement.