The AI Copyright Battle Begins  ## Metadata - Author: [[Hard Fork AI]] - Full Title: The AI Copyright Battle Begins - Category: #podcasts - URL: https://share.snipd.com/episode/fa1ca3ae-9dd3-4672-ad65-c04c55365064 ## Document Notes ## Highlights - **Historic $1.5B Copyright Settlement** - Anthropic agreed to a $1.5 billion settlement with writers over copyright claims tied to training data. - This marks the largest payout in US copyright history and sets a high-profile precedent. Transcript: Jaeden Schafer Anthropic has just reached a historic $1.5 billion settlement with writers for copyright. Now, this is basically a lot of people are excited about this, and there's also a lot of people that are not excited about this. Personally, I'll fall in the camp that is happy with the direction of the settlement and basically the verdict of it. I'll break down both sides, but there's definitely people that are unhappy. If you go over to TechCrunch, they a whole article that says, screw the money. Anthropics, $1.5 billion copyright settlement sucks for writers. So I'll put it out there. Not everyone's happy about this, but for AI companies, this is, you know, positive for the industry. So let's dive into all of it. Before we do, I want to say, if you want to try any of these models I talk about on the show, including all of the Claude models from Anthropics, love for you to go check out my own startup, Which is AIbox.ai. We have the top 40 different AI models all in one place. So you pay 20 bucks a month and you get access to 40 different AI models. So you don't have to have subscriptions to everything. But in addition, we just launched our AI app builder, which is basically a little box like chat GPT, and you type in a tool that you want to create, and it will chain together multiple AI Models, you will put prompts in, and it will basically build you a tool, you can go and customize it. We're really excited about this. And this is what we've really been working on for the last two and a half years. So if you want to go try out the no code AI app builder on AI box, there's a link in the description. And I would love to hear your thoughts as we're actively fixing things, adding things. And it's an exciting time for us over here. All right, let's get into the episode. So basically, if you go over to TechCrunch, like I mentioned, they're not super excited about this. But how this is rolling out is that about 500,000 writers are going to be eligible for about three thousand dollars in one point in this 1.5 billion dollar settlement a group of writers Brought this lawsuit against anthropic it's kind of interesting because it's not just the anthropic trained off of their data and that's kind of what a lot of people are complaining About with this lawsuit is how that was how that is kind of the shakeout on that but basically this is the largest payout in the history of US copyright law. And it is, I think, really exciting. ([Time 0:00:00](https://share.snipd.com/snip/b6b77621-e510-4563-b955-5e1bb83cb3ea)) - **Host's AI Toolbox And App Builder** - Jaeden plugs his startup AIbox.ai which aggregates 40 AI models and offers a no-code app builder. - He invites listeners to try tools like Claude via that platform and provide feedback. Transcript: Jaeden Schafer We're really excited about this. And this is what we've really been working on for the last two and a half years. So if you want to go try out the no code AI app builder on AI box, there's a link in the description. And I would love to hear your thoughts as we're actively fixing things, adding things. And it's an exciting time for us over here. All right, let's get into ([Time 0:01:11](https://share.snipd.com/snip/552e7f5b-52a9-4ce4-af80-39cb879ac7c0)) - **Shadow Libraries Fueled Claude's Writing Skills** - Jaeden explains Anthropic ingested millions of pirated books from 'shadow libraries' to rapidly improve Claude's writing tone. - They then purchased and scanned many books to cover their tracks and include them in training datasets. Transcript: Jaeden Schafer So for me, anyways, but some people do not think this is a win for authors. It is just a win for it is by TechCrunch. It is quote, yet another win for tech companies. Everyone is trying to get as much data as you possibly can to train the models. I think we all know this. Everyone basically scraped the internet at the very beginning. OpenAI scraped the whole internet at the beginning. Everyone did. And then people complained about that. Oh, you scraped the blog post. So that was kind of like a thing. What actually ended up happening is these AI model companies ran out of data. They wanted more data. They ran out. Like, they scraped the internet. And so a really interesting untapped source was books, right? Because books a lot of times are not actually online. Google has kind of their Google Scholar, I believe, a kind of project that has like photocopied books and put all the pages on. I don't think they allow people access to basically use that. And there's not all books on there. There's a ton that are not. So what Anthropic ended up doing was, and this is what got them in trouble, is they went to a bunch of pirated sources. They're called, quote unquote, shadow libraries. So there's millions of books in there and they're pirated sources, but it's not just a photocopy. People have uploaded the full book. There's the transcripts you copy and paste, right? So it's very easy data for these AI models to ingest. And it's virtually impossible to have gotten that data set as fast as it did. So they were able to grab all of these pirated libraries, throw them into the model, and Claude got way better. I think this is one of the reasons why, even compared to OpenAI from the early days, Claude has always had a much better tone in how it talks and writes. It's been way better for writing, basically. It's kind of ironic. But all the writers I know use Claude because, like, yeah, the tone's way better. And that's because they grabbed a copy, a pirated copy of every single book. Now, I think they kind of knew they were in hot water with this. They were in trouble. Maybe they're trying to cover their tracks or cover their back. And what they ended up doing was going and buying one of like every book in the world, like something crazy, right? And when you have billions of dollars, it's just a cost of doing business. Then they basically had a robot that would take each of these books, would flip through the pages, scan the pages, and then transcribe the pages. And then like basically take a picture and they can read the picture and then include that into the model data training. They did both of those things. ([Time 0:02:09](https://share.snipd.com/snip/5282218d-f3ce-4e32-b28e-73a1dac47158)) - **Judge Framed Training As Knowledge Acquisition** - The court treated scanning purchased books for model training as akin to a person reading to gain knowledge, invoking fair use. - Pirated sources remain illegal, but bought-and-scanned copies were ruled permissible for training under that framing. Transcript: Jaeden Schafer And when the lawsuit came, they got in trouble for obviously the millions of books on a pirated library. And this is actually what the billion dollar settlement is coming from now the judge actually ruled in this case that you know what they were doing where they were scanning all of the Books and uploading them and like basically they purchased the book and then they include it in the data set they said that is allowed because it's the same as if a person goes and buys The book and reads the book has the knowledge and they go write like some sort of paper or some sort of essay on it and that's monetized like you're allowed to do that because you gained Knowledge. And so this is kind of like what they did. They paid for the book and they gained knowledge. So you're not allowed to use pirated books, but if you buy the book, you can include it in your data set. And so a lot of people are upset because they're like, you know, those authors should have reoccurring compensation forever. If you want, if they want to be included, they should be able to be pulled in and out. I think the cat's out of the bag. It's kind of too late. Honestly, with the shadow libraries, it's too late now anyways, because basically if you have the pirated copies in the model, you could just use the old model to train a new model. And so even if you're like, okay, we're not using the old model anymore. The data is already in there. The tone's already in there. It's kind of too late at this point. ([Time 0:04:13](https://share.snipd.com/snip/102b51ca-8164-4d6d-af64-42a012b4540b)) - **Settlement May Set Legal Precedent** - The settlement resolves legacy claims but doesn't halt other pending suits against AI firms over copyrighted training data. - The Barts v. Anthropic outcome may become a reference point for future litigation about training on copyrighted works. Transcript: Jaeden Schafer And so it's now just like, what's their fine? So the fine was $1.5 billion. What's interesting is there is dozens of lawsuits filed against companies like Meta, Google, OpenAI, and MidJourney over basically all the legalities of training AI on copyrighted Work. So this isn't the first, I don't think this is going to be the last copyright one that we see come out. I think Anthropic is going to come out ahead for this. And I think a lot of people are happy with the precedent because now they know like the right way they can do this. I think everyone's kind of solved this problem at this point, but it's nice to know that for them anyways, that the way they've solved it is something that they can continue to do into The future and they're not going to get in trouble for. So right now, writers are basically getting a settlement if their work was included in all of the pirated stuff. So what's interesting is Anthropoc actually just raised $13 billion. I did a podcast on that if you're interested on the dynamics of that. But they've just raised $13 billion. So paying out 1.5 is not going to kill them, right? Their last raise was 3.5. And I imagine if they had to pay out 1.5 billion of 3.5 just for that one lawsuit, that would be hurting them quite a lot. I think with this fresh round of funding, they can move forward and they'll be fine. But all of this happened because in June, federal judge William Alsup sided with Anthropic and ruled that it is legal to train AI on copyrighted material. ([Time 0:05:21](https://share.snipd.com/snip/0fa74d62-41b5-4955-bd2d-c7319ca0dec1)) - **Avoid Pirated Data; Buy Or License Corpora** - AI companies should avoid using pirated materials and prefer licensed or purchased corpora for training. - Expect firms to adapt practices to rely on purchased datasets or negotiated licenses to reduce legal risk. Transcript: Jaeden Schafer He argued that this use case is transformative enough to be protected by the fair use doctrine that is carved out of copyright law that was set back in 1976. So he said, quote, like any reader aspiring to be a writer, Anthropics LLMs train on open work upon works, not to race ahead and replicate or supplement them, but to turn a hard corner And create something different. The piracy obviously was a completely different problem. And that's why he let the case go to trial was because of that. And this is what Anthropic said about this whole thing. They said, quote, today's settlement, if approved, will resolve the plaintiff's remaining legacy claims. This is Aparna Sridhar, who is their deputy general counsel at Anthropic. And then they also said, we remain committed to developing safe AI systems that help people and organizations extend their capabilities, advance scientific discovery, and solve Complex problems. So there are tons more cases that are currently being litigated right now between AI and copyrighted works. But because of this, I think this, you know, Barts versus Anthropic is going to be basically a precedent in all of these things. And so I think that there's some people think that because of the ramifications of it maybe judges are going to arrive at a different conclusion but i think basically this precedent Is going to hold and we're going to start to see that a lot of these cases i won't move forward if you're using pirated stuff obviously you're gonna get in trouble but if you purchased whatever The original work was and like you kind of think of like music generators which i i you know this would be funny and they'll probably all get slaps on the wrist too. But like, you'd have to go buy one copy of like, every top song ever recorded for the last 100 years. And then no, no, whatever a billion dollars you spend on that, you know, all ever all the music in the world, then you can feed that into your AI model. So to make music. So I think basically, we have a precedent for how it should go. ([Time 0:06:37](https://share.snipd.com/snip/b9c44d63-1975-487f-a474-98aece31a4b8)) - **Provenance Tracking Is Practically Impossible** - Tracking provenance of every training example is practically impossible, making per-use royalties hard to implement. - Models can perpetuate copyrighted style and content once data are absorbed, so perpetual tracking offers limited protection. Transcript: Jaeden Schafer In my opinion, I think this might be the way you have to go because one of the big problems that like Adobe tried to solve with image generation was they were like, they're like, look, we'll Pay people if you include your images in our data set, they pay the original photographers for Adobe Firefly images. But they it's impossible to know, like when I say, you know, generate a picture of a green plant on a stand with a flag in the background, like What data was used to create that image? What was needed? It's not like you could do it like Spotify, where if you listen to a song, they just listen to that song. So now you give them a penny in streaming revenue. It's impossible to know what the original source was. So basically, Adobe did it where they just took in a huge data set of images, and they're like, look, we took in a million images. So let's say there's a million photographers each put in one image, they all get like one, one millionth of the, of the royalty or revenue or whatever. I think it's, I don't think it's realistic to set up systems where like, once you're included in a data set, now all of a sudden you can use that model to spit out more outputs that other Models can use to train on it. It's just really, it's, it's lost. So I think it's impossible to track everyone's copyrighted data forever. ([Time 0:08:23](https://share.snipd.com/snip/b0391483-4989-4b92-bd8b-1b4e4b7e17e8))