3 ms·

> I firmly believe that training models qualifies as free use. I think it falls under research, and is used to push the scientific community forward. I don't t

by kopecs 3y ago

> I firmly believe that training models qualifies as free use. I think it falls under research, and is used to push the scientific community forward. I don't think this is as cut-and-dry and you make it out here. If I train a model on, say, every one of New York Times' and release it for free and it finds use as a way of circumventing their paywall I have difficulty justifying that as fair use/fair dealing. The purpose/character of the model should indeed be a factor but certainly not nearly as dispositive a one as I think you're suggesting. To the extent that training the model serves a research purpose I think that the general use / public release of the trained model does not in general serve the same research purpose and ought to have substantially lower protection on the basis of, e.g., the effect on the original work(s) in the market.

jacobr1 3y ago
Wouldn't that depend on the use case? If you just had the model regenerate articles that roughly approximate its source material that is much a more clear cut violation of a paywall. But if you use that data as general background knowledge to synthesize aggregative works such a history of the vietnam war, or trends in musical theatre in the 1980s relative the 1970s, or shifts in the language usage of formal honorifics, then that seems to me to be clearly fair-use categories. There are gray areas, such as aggregating the opinions of a certain op-ed writer over a short timeframe that while it might produce a novel work, is basically is mixmash of recent articles. But would that be unfair, especially if not done in the original authors style? These technical distinctions like these probably will matter in whatever form regulation eventually ends up becoming.
- kopecs 3y ago
  Yes, I think this is a rather fact-specific inquiry. My main point is that the research/commercial distinction is not the only factor (and not even the most important one). > if you use that data as general background knowledge to synthesize aggregative works such a history of the vietnam war, or trends in musical theatre in the 1980s relative the 1970s, or shifts in the language usage of formal honorifics, then that seems to me to be clearly fair-use categories I don't think this is clear. If someone were to train a model on several books about the Vietnam War and then publish my own model-created book on history of the Vietnam War, I would be inclined to say that that is infringement. And if they changed the dataset to include a plurality of additional books which happen to not be about Vietnam, I don't think that changes the analysis substantially. I think it is hard to earnestly claim that in that instance the output (setting aside the model itself, which is also a consideration) is transformative, and so I would think, absent more specific facts, that all four fair use factors are against it.
  - jacobr1 3y ago
    > And if they changed the dataset to include a plurality of additional books which happen to not be about Vietnam, I don't think that changes the analysis substantially. I think the question is if it changes analysis if the dataset DOES include a bunch of books and articles related to Vietnam beyond your specific book. In the first cast where it just rewriting a single books content, the unfairness is clear. But in case where it is producing a new synthesis and analysis of the data, derived in part, but not regurgitating the source material, is that unfair? The latter isn't clear to me.
- makomk 3y ago
  Quite a lot of what news publications like the New York Times do is precisely regenerating articles that roughly approximate source material from some other publication. If I remember rightly, a lot of smaller, more local news organisations aren't happy about this because of course it's a more or less direct substitute for the original and a handful of big news organisations (particularly the New York Times) are taking so much of the money that people are willing to pay for news that it's affecting the viability of the rest - but it's not illegal, since facts cannot be copyrighted.

by kopecs 3y ago

jacobr1 3y ago

Wouldn't that depend on the use case? If you just had the model regenerate articles that roughly approximate its source material that is much a more clear cut violation of a paywall. But if you use that data as general background knowledge to synthesize aggregative works such a history of the vietnam war, or trends in musical theatre in the 1980s relative the 1970s, or shifts in the language usage of formal honorifics, then that seems to me to be clearly fair-use categories. There are gray areas, such as aggregating the opinions of a certain op-ed writer over a short timeframe that while it might produce a novel work, is basically is mixmash of recent articles. But would that be unfair, especially if not done in the original authors style? These technical distinctions like these probably will matter in whatever form regulation eventually ends up becoming.

kopecs 3y ago
Yes, I think this is a rather fact-specific inquiry. My main point is that the research/commercial distinction is not the only factor (and not even the most important one). > if you use that data as general background knowledge to synthesize aggregative works such a history of the vietnam war, or trends in musical theatre in the 1980s relative the 1970s, or shifts in the language usage of formal honorifics, then that seems to me to be clearly fair-use categories I don't think this is clear. If someone were to train a model on several books about the Vietnam War and then publish my own model-created book on history of the Vietnam War, I would be inclined to say that that is infringement. And if they changed the dataset to include a plurality of additional books which happen to not be about Vietnam, I don't think that changes the analysis substantially. I think it is hard to earnestly claim that in that instance the output (setting aside the model itself, which is also a consideration) is transformative, and so I would think, absent more specific facts, that all four fair use factors are against it.
- jacobr1 3y ago
  > And if they changed the dataset to include a plurality of additional books which happen to not be about Vietnam, I don't think that changes the analysis substantially. I think the question is if it changes analysis if the dataset DOES include a bunch of books and articles related to Vietnam beyond your specific book. In the first cast where it just rewriting a single books content, the unfairness is clear. But in case where it is producing a new synthesis and analysis of the data, derived in part, but not regurgitating the source material, is that unfair? The latter isn't clear to me.
makomk 3y ago
Quite a lot of what news publications like the New York Times do is precisely regenerating articles that roughly approximate source material from some other publication. If I remember rightly, a lot of smaller, more local news organisations aren't happy about this because of course it's a more or less direct substitute for the original and a handful of big news organisations (particularly the New York Times) are taking so much of the money that people are willing to pay for news that it's affecting the viability of the rest - but it's not illegal, since facts cannot be copyrighted.