> In the AI use case, they're typically aiming not to output any significant part of the training data
What they’ve aimed to do and what they’ve done are two different things. Models absolutely have produced output that closely mirrors data they were trained on.
> not competing in the market with the original work
This seems like a stretch, if only because I already see how much LLMs have changed my own behavior.
These models exist because of that data, and directly compete by making it unnecessary to seek out the original information to begin with.
And yet, the IA is 100% aiming to absolutely reproduce literally every part of the work in a 100% complete manner that replaces the original use of the work.
And you cannot bring yourself to admit that the IA is wrong. When you get to that point you have to admit to yourself that you're not making an argument your pushing a dogma.
The point more generally is that there’s an asymmetry in how people are thinking about these issues, and to highlight that asymmetry.
If it turns out after various lawsuits shake out that LLMs as they currently exist are actually entirely legal, there’s a case to be made that the criteria for establishing fair use is quite broken. In a world where the IA gets in legal trouble for interpreting existing rules too broadly, it seems entirely unjust that LLM companies would get off scott free for doing something arguably far worse from some perspectives.
What publishers argue is that you cannot treat digital books like physical ones; i.e. you cannot re-sell or lend (like IA did) a digital book.
What LLM do is that they use copyrighted content for profit and do not lend anything.