I don’t necessarily fault OpenAI’s decision to initially train their models without entering into licensing agreements - they probably wouldn’t exist and the generative AI revolution may never have happened if they put the horse before the cart. I do think they should quickly course correct at this point and accept the fact that they clearly owe something to the creators of content they are consuming. If they don’t, they are setting themselves up for a bigger loss down the road and leaving the door open for a more established competitor (Google) to do it the right way.
I agree. You can even listen to the NYT Hard Fork podcast (that I recommend btw https://www.nytimes.com/2023/11/03/podcasts/hard-fork-execut...) where they recently had Harvard copyright law professor Rebecca Tushnet on as a guest.
They asked her about the issue of copyrighted training data. Her response was:
""" Google, for example, with the book project, doesn’t give you the full text and is very careful about not giving you the full text. And the court said that the snippet production, which helps people figure out what the book is about but doesn’t substitute for the book, is a fair use.
So the idea of ingesting large amounts of existing works, and then doing something new with them, I think, is reasonably well established. The question is, of course, whether we think that there’s something uniquely different about LLMs that justifies treating them differently. """
Now for my take: Proving that OpenAI trained on NYT articles is not sufficient IMO. They would need to prove that OpenAI is providing a substitutable good via verbatim copying, which I don't think you can easily prove. It takes a lot of prompt engineering and luck to pull out any verbatim articles. It's well-established that LLMs screw up even well-known facts. It's quite hard to accurately pull out the training data verbatim.