Speaker 0:  acort has effectively handed the ai industry a $1,500,000,000 receipt for bad data sourcing. anthropic's punishment was not simply for teaching an ai with books. it was for allegedly building a private library from millions of pirated copies. so silicon valley's first download first, ask permissions later era just received a $1,500,000,000 reality check to court, left the door open for ai training, but put a very expensive lock on a pirate entrance. so ai camp companies may still be allowed to study the books, but apparently, they cannot arrive at the library wearing an ipad and carrying a 7,000,000 stolen pdfs. anthropic settlement now puts a billion dollar price tag on the providence. so, basically, they didn't have a problem with them using the content. they're having problem with them where they found the content, where they, like, from pirate bay and other places, grabbed it, stored it locally, and then, in a private database server that no one else could access and then had their servers pour through it. and it wasn't that they looked at the books. it was where they sourced it and how they used them, specifically, you know, illegally that is the problem. what do you think about this, john?

Speaker 1:  yeah. so one of the biggest misconceptions about this case is that it determined whether ai training itself is illegal, which it did not. okay? so there were two questions. well, the court separated the dispute into two independent legal questions. the first question was, was acquiring sorry. i laugh at this question, though. was acquiring millions of pirated books legal? no. obviously. so, and throw like you said, anthropic allegedly downloaded millions of copyrighted books from well known piracy repositories, which violated copyright law. right? yes. it did. is the second question that was part of this case was, is training an ai model using books transformative enough to qualify as fair use. and so the earlier ruling by judge william alsop concluded that under the facts before the court, the training process itself could qualify as transformative fair use, and that finding remains intact. so if they had actually bought the books and paid for the books, this there wouldn't be a problem here. so the settlement concerns the unlawful acquisition, jesus, of storage and copyright works, not the abstract legally of machine learning itself, which i think is really important as well here. so my question is does this open like, is is anthropic the only company that did this? how did they find out that anthropic did this, i guess, through discovery? and does this open up for openai and all these other llms to be sued for the same thing?

Speaker 0:  yeah. this does, give ai companies, publishers, investors, a measurable cost for questionable training data sourcing. so because this 1,500,000,000 settlement, it covers, from what i understand, 486,000 408 sorry. 482,460 books acquired from pirate libraries. so there is now a potential precedent depending on whether things get appealed or don't. ai training survived, but pirate sourcing did not. so both sides can legitimately claim a partial victory. anthropic preserved an important fair use ruling covering copies made during training. author secured the largest known copyright settlement and compensation for mass downloading from private repositories. but the decision, like you said, does not establish that every ai training use is legal. it does not establish that every use of copyrighted material requires a license. so the consequences of becoming, data provenance is where the material came from, which copies were made, and why were they retained, and what rights the developer had at each stage. so, you're right about, you know, what it covers and what it does cover. but this is a warning that training data is becoming a governed corporate asset rather than an anonymous pile of internet content. the settlement gives publishers and rights holders a negotiated

Speaker 1:  yeah yeah. exactly. i know i got to say i'm surprised by anthropic because they're the ones that normally will are seen as the good guys here. so, but, again, when they did this, this was in the heyday of, like, you know, we need to train our models. we need to get this done. what is the cheapest way to do this? and we'll just ask for forgiveness after, and that's kind of what they did. and that's what a lot of these ai companies have done and are still doing. they're still gonna keep this really is not gonna stop these companies to keep doing this, i don't think, because they need to train their models. and so they're gonna do it by any by the cheapest means possible for now, until they get caught. and they're gonna keep doing it in my opinion. what do you think? do you think this is gonna stop ai companies from doing methods like this or using methods like this in the future?

Speaker 0:  well, when we say methods like this, basically, the problem with all of this is not that they used it, but that they got it from pirate bay or other, illegal sources, and then that they held on to these illegally sourced machines in some sort of central repository for continued use. so those two things were the big issue here. so i guess if you source it right, why not? and i think at some point, it's gonna be who these companies just to buy the books or buy the sources. right? $3,000 a book is pretty expensive when you could have bought the ebook for $20 probably. right? so, you know, at some point, they're just gonna buy it and say, okay. we bought it. we have fair use. with it now, we can do whatever we want with it based on fair use rules, and the writer gets their fair share. now i think we're probably moving into a now that we understand this part, let's get into what does licensing my book to look like? okay. you're gonna use my book. i want $1,500. you're gonna use my book. i want 3,000, not $20. right? so i think we're at a negotiating publisher point where, okay, we have the $3,000 per illegal use of my book kind of situation. so that is a stake in the sand that we can now negotiate against for future rights. so i think we're gonna see this as a stake in the ground starting point, and it's about time. i mean, you know, talk to air. we haven't seen much of that stuff come out yet as to the results of some of those lawsuits, so we're gonna see some of them come out soon. it'll be interesting to see, you know, what is considered fair use and what isn't, what is considered a proper source, and what isn't, you know, in all of these things. i'm curious to see, but i think this is not a shot across the bow because the lawsuits are what's needed to put the stake in the ground. you know, we kinda differ on some of these things around, you know, the process to get to the rules, and it's okay as long as we get to the rules. and i think this is the first step to get to those rules.

Speaker 1:  i agree. now there's some open questions that remain. right? so this case itself, it leaves several unresolved issues. does fair use apply equally to images, music, video, news, computer? like, is it just books? there's so many questions that still are are there is so much litigation that is going to happen in our lifetime related to ai and so much precedent and so much, so many decisions are gonna be made that it's going to be nuts in my opinion.

Speaker 0:  yeah. i think you're right that it is gonna be nuts. and the question is, are they going to go with the least restrictive or the most restrictive? i say that for example, we have a us ruling. what about a eu ruling? what about a china ruling? what about a canada ruling? now they all have slightly different rules, slightly different prices, bear where you can source it, what fair use means. are these companies gonna go with the most restrictive to be compliant, or are they gonna go with least restrictive or somewhere in between? that becomes the question as these things go in place. you know? and regulation eventually has to happen. right? because you can't just use everybody else's information all the time because where does the creative activity stop then? we're not gonna create anymore and think through anymore stuff. we're just regurgitating old things. and, you know, where does copyright begin? you know, companies like china doesn't really believe in copyright. the state owns everything, from a copyright intellectual property point of view where us is a firm believer in, hey. if you create something new and can get some ip around it, then you deserve to make money off it for a period of time. whether that's the right period of time and all that kind of stuff, that's debatable, but that's their stance. so yeah. the question is who, you know, in the end, this is gonna get appealed, and we're gonna get some stuff around it. it might not get appealed because it's not too bad from a point of view. but some legislation needs to come out of it, and we'll have to see what legislation hits the different countries and how that works out to, the us, the consumer, who wants to use these things. right? is it gonna be able to get quotes for us that we want anymore? is it gonna be able to reference, you know, if i'm doing a paper on literature and i say, go get me three quotes from three famous, is it gonna be able to do that? or is it gonna say, yeah. you know what? you're i, as a user, are gonna have to pay a fee for that as part of my token thing? or is it something, like, the world's changing. same thing with, hey. create images in this style. you know, they have the pixar style, and they have that i can't remember what that style the gimli style that got very popular for a while. yeah. yeah. yeah. so all those are not a 100% copyrightable, but they seem to be challenged and courts will have to see what happens. so all of this stuff is at question. and i think we've crossed the line saying, hey. there is a fair use in using this stuff. it becomes question now of reusability, what the output matches the input, and where you source this from legally or not. now does that relate to artist style and whatnot rather than information from writing? we'll have to see. this is a very landmark case, and, i'm interested in seeing like, it looks like, it's all handed out. so the money has almost all been handed out already, so there's not gonna be much in a bit appeal like this. so think of this like a chef learning recipes from cookbooks. right? when the ai is learning all this stuff, is that really a valid output? so, the settlement doesn't does not allow ai training, but it makes one thing painfully clear. transformative technologies does not automatically sanitize an unlawfully sourced dataset. the next ai arms race may not be about who has the most data. it may be about who can prove where their data came from. would you trust an ai company more if it published a fully nutrition label showing exactly what trained its models. let us know your thoughts in the chat below.