If it truely were impossible, then it sounds like they've just admitted that they should have licensed the content rather than using it without permission.
OpenAI's quote, from which the headline is derived, briefly mentions public domain content:
> > Because copyright today covers virtually every sort of human expression including blog posts, photographs, forum posts, scraps of software code, and government documents – it would be impossible to train today’s leading AI models without using copyrighted materials. Limiting training data to public domain books and drawings created more than a century ago might yield an interesting experiment, but would not provide AI systems that meet the needs of today’s citizens.
I do think they're right in that you're not going to get remotely the same performance as modern LLMs if you train with solely public domain content. Licensing a sufficient amount of content could be feasible for OpenAI/Meta, but hopefully we get something that still works for smaller companies, FOSS groups, etc.
Even being in favor of massive copyright reform, I don't think I can get behind that argument.
My intuition is that, if you publish something online and let me read it, I can use that knowledge. I can NOT flat out duplicate it for others, but I can still use the knowledge. Training OpenAI is akin to me using something I read. (Hell, most of what I know I obtained this way.)
To me, the only issue is if the model is reproducing text verbatim en masse.
I think my intuition above is basically where my intuition lands for search engines too. The can use datas found on the web, and can show me snippets, but I'd draw a line at them showing me all/most of an article.
I think that's a subtly different angle since it doesn't make any appeal to scale/cost. Sort of like the difference between:
1: "Everybody should be free to use Article X in manner Y without negotiation or payment, as a matter of principle."
2: "When I choose to use millions of Article X'es in manner Y, it just becomes expensive and impractical to negotiate/pay, therefore my situation is qualitatively different and exempt."
Think of my intuition as an all you can eat buffet. In spirit, you can only eat so much as one person and you may bring home some scraps of food you already touched. In theory (if we ignore fine print), you can very much eat out half the buffet, and then take the other half to go. But this latter "approach" breaks down the model of a buffet. Tragedy of the commons. Everyone is a glutton, the buffet can't keep up with making food, business shuts down, no more food for anyone.
The internet here is a buffet. Just because you have access to terabytes, petabytes or more of information without paying doesn't mean you can download all those terabytes of info. Everyone starts doing that and the model of the commons that are ad-supported servers breaks down. Sites start to close up behind more anti-copy protection or literal paywalls to make things not-so-freely accessible. And that ends the free internet as we know it. We are indeed starting to see pieces of this happen in real time.
>To me, the only issue is if the model is reproducing text verbatim en masse.
That's the other issue as well that was discovered with that exploit. You can reguritate the source with enough prompting. This seems to imply that some server out there is in fact just storing hoards of source material on itself. If that server is commercialized and has TBs of copyright material, that seems to go past the fair use argument of "limited use".
Sites have been pushing towards "to continue reading, download our app" or "you've reached your free article limit, subscribe for just $X" for a long time, because it's profitable. I don't think there's sense in retroactively placing the blame for this behavior on crawlers - it's a tactic designed for humans, to the extent that it's often implemented in ways entirely ineffective against bots such as putting an element in front of the content, using cookies to track viewed article count, or even intentionally letting bot user-agents through the paywall. Sites are capable of restricting GPTBot/CCBot/etc., or putting a rate limit on crawling, without impacting human readers.
> You can reguritate the source with enough prompting. This seems to imply that some server out there is in fact just storing hoards of source material on itself. If that server is commercialized and has TBs of copyright material, that seems to go past the fair use argument of "limited use".
The model isn't retrieving content from some storage server during inference if that's what you're thinking, but consider Google Books which does actually do that - storing a huge number of in-copyright books internally on Google's servers and retrieving snippets when searched. As found in Authors Guild v. Google, this is Fair Use - in particular:
> > What matters in such cases is not so much "the amount and substantiality of the portion used" in making a copy, but rather the amount and substantiality of what is thereby made accessible to a public for which it may serve as a competing substitute.
And websites have been getting paid for a long time by selling user data (which again is only useful for human data, not robot data). I don't think it's a coincidence that the paywalling of more and more sites came as GDPR and other data privacy regulations started to pop up. Crawlers are one of those smaller symptoms that gets knocked out along the way as well. Can't make money one way, they switch to another (even before taking the pandemic and interests rates into account).
>Sites are capable of restricting GPTBot/CCBot/etc., or putting a rate limit on crawling, without impacting human readers.
all those captchas I'm being hit with these days tells me otherwise. Maybe it is possible, but perhaps too expensive. Or the capthas themselves are great for training data. I don't know the real reason, but the internet has definitely become most hostile towards my human eyes in an attempt to curb bots. And the worst part is that it doesn't even entirely block out the bots.
>The model isn't retrieving content from some storage server during inference if that's what you're thinking
I don't admit to know the full details, but some exploits discovered lately seem to suggest that this isn't just "reading a page" and then moving on: https://arxiv.org/pdf/2301.13188.pdf (PDF warning). Apparently somewhere in the memory there is those raw assets being stored for "reference".
Again, I don't know enough to debate on it, but I've seen enough pieces to be scrutinous of how the process currently works. Scraping is a gray area, but storing copyright data in your commercial product is a much harder issue to argue. Youtube screws its creators over backwards to prevent litigation over a very similar concept.
Data protection regulations being the cause of paywalls seems a different theory than what you were previously suggesting.
> all those captchas I'm being hit with these days tells me otherwise. Maybe it is possible, but perhaps too expensive.
Captchas are generally to prevent spambots. CCBot/GPTBot crawlers identify themselves by their user-agent (so the server can just choose not to serve the content) and also respect robots.txt.
> https://arxiv.org/pdf/2301.13188.pdf (PDF warning). Apparently somewhere in the memory there is those raw assets being stored for "reference".
Stable Diffusion is ~4GB and trained on 5 billion images - for the most part it's not plausible for it to have stored the "raw assets". There is a real issue of overfitting on a subset of content that the model saw many times during training such that it was able to memorize details, but the model isn't retrieving the source material from some server to reference (that's known as retrieval augmented generation, which isn't used by default SD/GPT).
Would also keep in mind that for papers like this, where they're generating hundreds of millions of images and digging into CLIP embeddings to construct prompts that intentionally recreate images, that there's some degree of similarity you'd expect just by the conditioning given, the large number of comparisons being automatically made, and coincidence - even had the model not seen the image in question.
> Youtube screws its creators over backwards to prevent litigation over a very similar concept.
Youtube makes the video available to the public. The Google Books analogy properly preserves the aspect that full works are stored but only small portions made available.
Sure. I guess we'll see how that holds up in the lawsuit over the coming years.
> for the most part it's not plausible for it to have stored the "raw assets"
I wasn't talking as literal pngs on a server. I'm sure a company backed by Microsoft has much more clever ways to store such data. My main point was that it seems possible to reproduce the asset so enough of it is kept to perform such operations.
>Would also keep in mind that for papers like this, where they're generating hundreds of millions of images and digging into CLIP embeddings to construct prompts that intentionally recreate images, that there's some degree of similarity you'd expect just by the conditioning given, the large number of comparisons being automatically made, and coincidence - even had the model not seen the image in question.
coincidence, perhaps. I can't say and tracing isn't technically illegal either (although extremely faux pas for artists doing anything more than learning how to draw). But that is definitely where the metaphor of human motor skills and machine reproduction breaks down. Humans can more easily be cast off as coincidence based on context (which can be easy or hard to deduce. E. G. Did artist B follow artist A on Twitter? How long were they active? Are they known to have a similar style?)
bots' context is closed up as company IP. That alone will give it a lot more scrutiny. Scrutiny that can potentiallybe subpeona's, so again: TBD.
>Youtube makes the video available to the public. The Google Books analogy properly preserves the aspect that full works are stored but only small portions made available.
I don't really know why you're so fixated on the Google books metaphor. Google books is also an actual bookstore as well as some place for users to upload their library. I can upload pirated content for private use becsuse Google isn't scanning all its customers 24/7. It's technically not legal but also hard to enforce on my private library of 1 that I share with no one.
I'd run into more issues if I tried to sell it and someone caught me. You can DMCA a book from Google just as easily as a YouTube video. There's simply less money in the game to DMCA troll literature (and for many books of interest to HN, the primary customer isn't individuals but institutions and education).
It's an example we have a court ruling for and is a more literal case of internally storing a huge number of copyrighted works, without permission of the copyright holders, then allowing users to query for snippets.
> Google books is also an actual bookstore as well as some place for users to upload their library. I can upload pirated content for private use becsuse Google isn't scanning all its customers 24/7.
Google Books is a search tool for millions of books that Google have scanned and OCR'd. As far as I'm aware it doesn't itself sell books and I don't believe there's a way for regular users to upload content.
Are you thinking of books on Google Play?
> It's technically not legal [...] You can DMCA a book from Google just as easily as a YouTube video
Google Books has been found to be legal due to Fair Use. You could send a DMCA takedown notice, but they wouldn't have to comply with it.
AI's gonna get trained on it no matter what; nobody wants a product that isn't. By the time OpenAI takes the bullet we will have killed nothing.
If you can convince adverts that a scaped hit counts as an organic view, I can concede on your point of "educating yourself on copyrighted material is okay". But given NYT here, that seems to already have been a lost point.
>AI's gonna get trained on it no matter what; nobody wants a product that isn't.
That's fine. Just license it like everyone else that wants to use copyrighted material for commercial use. Why is that so hard? It's not like Microsoft can't bankroll it.
It's true but still breaking the law.
But I guess the sad fact of it all is that I'm sure they 100% anticipated lawsuits, and expected that fighting them is still cheaper than trying to license all the content properly. I doubt anyone is surprised by this lawsuit. Just disappointed that they will throw money at anyone other than the actual artists and writers.