Open AI Is In SERIOUS Trouble (Open AI Lawsuit Details)

So this video will discuss one of the most important lawsuits of all time, the OpenAI versus the New York Times. So if you haven't heard, the OpenAI company, in conjunction with Microsoft, is being sued by the New York Times for alleged copyright infringement. And the New York Times are claiming that OpenAI trained its GPT4 model on many different New York Times articles and they essentially want GPT4 to be shut down. So pay attention to this because if this case is ruled in a certain way, it, it would have serious ramifications for the generative AI industry. And with that being said, let's get into exactly why this is so important. So let's take a look at exactly what's happening here, because essentially, if you don't know the New York Times, a media establishment is essentially saying that OpenAI has trained GPT4 and GPT 3.5, but mainly GPT4 on proprietary data from the New York Times, such as their articles, and they're essentially claiming copyright infringement and copyright theft. Okay, so that is the case. Now what you're seeing on screen here, as you can see, a series, well, this is just one, but what you're about to see, because I'm about to show you a lot, is you're about to see a series of essentially prompts that people have prompted ChatGPT with GPT4. And then the output, and then on the right hand side you can see that the actual output, the actual text from the New York Times and, and how similar it is. So you can see right here that this is the text, it says output from GPT4, a documentary film that premiered on a Wednesday, significant outbreak from his predecessors and the actual text from the New York Times. Now you don't need to read all of this, but the only text that isn't different is the black text. So the word premiered here, but all the red text is essentially the exact same. So there's a lot to discuss in this because I think this is one of the most important things because this kind of AI generative AI lawsuit is definitely one of the first of its kind. And we don't know how to try this because the laws haven't been updated for this kind of technology. So that's why this is really important. So you can see here why the New York Times is essentially saying that, look, OpenAI stole trained the data on our articles and when we've got this output, when we've prompted it, we've got the same exact stuff. But let's actually take a look further because someone actually did a Twitter thread. So I'm going to go on Twitter through this Twitter thread, you know, just talk about certain things and then essentially I'm going to show you some things that I did find when looking through the paper, because the paper, although it is long, I've read most of it and it's pretty, pretty crazy. So you can see here that it is basically a landmark case. And he said it's rooted in copyright law and the U.S. constitution, and that's very much where it begins. So it says the Constitution and the Copyright act recognize the critical importance of giving creators exclusive rights over their works. They since our nation's founding, strong copyright protection has empowered those who gather and report news to secure the future of their labor and investments. And it says copyright law protects the Times expressive original journalism, including but not limited to its millions of articles that have registered copyrights. And it says, Defendants, this is OpenAI have refused to recognize this protection. Powered by large language models containing copies of Times content, Defendants generative AI tools can output that recites Times content verbatim, closely, summarizes it, and mimics its expressive style, as demonstrated by the scores of examples see Exhibit J. Now, if you do take a look at Exhibit J, I'm not going to lie to you guys, there are a ton of different examples there. So I want to show you guys Exhibit J because they're like, hey, take a look at Exhibit J because this is pretty crazy. So this is Exhibit J here, and you can see that it says 100 examples of GPT4 memorizing content from the New York Times says the following are examples of situations where OpenAI's GPT4 model was trained on and memorized articles from the New York Times. Each example focuses on a single news article. Examples were produced by breaking the article into two parts. The first part of the article is given to GPT4. GPT4 replies by writing its own version remainder of the article. In each case, we observe that the output of GPT4 contains large spans that are identical to to the actual text of the article from the New York Times. For each example, we provide the following the URL, the prompt that was given, the response from GPT4, and the end of the article as it appears on the newyorktimes.com so the text is depicted in the red font when it appears identically in both GPT4 and the output and the source article from the New York Times. Now, when we take a look into this, you can see that essentially the output is pretty similar. I mean, for this text, this is all exactly the same because all the red text is here. And I'm only going to go over a few of these, but I'm scrolling down and you are seeing that there are literally over 100 different examples of this. Okay. And this case is so much bigger than people think because there's a lot of stuff that's going on on OpenAI, which means it just isn't really good for them. So this kind of lawsuit that people might be thinking OpenAI is a billion dollar company, this is completely fine. But as we do a deep dive into this case, you're going to see that, that this is really, really, really important because it sets the precedent on what exactly is going to happen. And I mean, you can see that like as I scroll down, as I scroll down, there is tons and tons and tons of this. But anyways, if we go back to this, we can see here that it then goes on to state that, but it does well to look towards the future and showing how violation used to create a substitute undermining existing and future business models, including AI licensing, which fund the critical and costly journalism around the globe. And so here they talk about how Microsoft is becoming a huge company based on the fact that they were able to acquire OpenAI, buy half of them. So it says Microsoft's deployment of the Times trained large language models through its product line helped boost its market capitalization by a trillion dollars in the past year alone. And OpenAI's release of ChatGPT has driven its valuation as high as $90 billion dollars. So I mean, what they're saying here is, look, you use the articles, you train the model on it, and because your model is so effective, you've essentially managed to get this huge sum of money. Like there's huge sums of money at stake here. Like they're a $90 billion company and that is of course, a huge amount of money. Now, essentially, they say the Times objected after it discovered that the defendants were using and anytime I do say the defendants, defendants are just OpenAI Microsoft and it says defense defendants were using Times content without permission to develop their models and tools. For months, the Times had attempted to reach a negotiated agreement with the defendants.

In accordance with its history of working.

Productively with large technology platforms to permit the use of its content in new digital products, including news products developed by Google, Meta and Apple. The Times goal during these negotiations was to ensure it received fair value of the use of its content, facilitate the continuation of a healthy news ecosystem and and help develop generative AI technology in A responsible way that benefits society and supports well informed public. So essentially what the Times are saying here is that, look, we discovered you were using this and we tried to reach out several times but we weren't able to reach a deal, which is pretty unfortunate because I think OpenAI did reach another deal with another other companies. And then here's where essentially they talk about the cost because it says this series was a product of an enormous amount of effort across three continents. And reporting this story was especially challenging because Times repeatedly denied was repeatedly denied both interviews and access. And basically they're saying that certain articles that they did were really, really hard to do and they had to go through so much trouble to get those interviews. And then essentially they're saying that they basically ripped that off and they're saying that these companies aren't really valuing how much stuff goes into the journalism of these stories. And then of course they're talking about how they provided real time detailed information across a range of topics, including US Elections. Many things that went around that I don't want to say because I don't want the video to get demonetized, but basically tons of stuff and real time information that a lot of people did use. Now what's also interesting is that they said and with clearly tied role for Microsoft, the complaint highlights the abuses that even in the most recent months it shows this example of where content was lifted verbatim from a New York Times report and then it compares to the approach taken by a search engine. So basically what they're saying here is that what we have is we have someone asking Bing Chat something. And then essentially what Bing Chat did was Bing Chat allegedly took the entire tree of a New York Times article and basically put it into that chatbot.

And of course this isn't good because.

It means that people don't need to find an article, it just needs to give that information with you in that chat session. And the problem with that is that people don't need to go to New York Times anymore. And they're alleging that secret. Here you can see New York Times article that they're alleging here that that is infringement because they're taking the content and putting it somewhere else. And then here's when they show another example of Microsoft, Bing, essentially just taking New York Times articles. And what's also interesting is that they claim that while OpenAI, Microsoft engaged in wow scale copying from many sources. They gave the Times content particular emphasis when building their large language models, revealing a preference that recognizes the value of their work. So they're basically saying that, you know, they did copy a lot of people, but they essentially prioritized our work because our work was probably the best work ever. They also talk about how the Times content was a total of 200,000 URLs accounting for 1% of all sources listed in the open web text 2 an open source recreation of the web 2 TechSetsu used in training GPT 3. So of course on the slide we do have more and more different articles because like I said, there are, you know, several, you know, examples that are here. And then what I do find interesting is that they say that despite its early promises of altruism, OpenAI quickly became a multi billion dollar company for profits. Yada yada yada. So basically what they're saying here is that OpenAI, firstly, if you didn't know the story of OpenAI and why they're called OpenAI is because essentially OpenAI was meant to be an open source AI company, which means that they were supposed to do open research, they were supposed to share what they did and over time they've lost their way and they've become a for profit company that is closed source, that don't have have any open source code. So people, and now the New York Times are saying that now they became a billion dollar company for profit built on the unlicensed exploitation of copyrighted works that belongs to the Times and others. And essentially they turned off their nonprofit status in around March 2019 and now they're just basically a, they're basically not a conglomerate of course, but mainly just like a profit hungry company. Now this is why this is one of the most widely discussed things on the Internet is because one thing that they do say, and I don' think that this is going to happen, but they do say that, you know, okay, so let's just take a look at some of the things that they do want. Okay, so it says we, you know, we want awarding the Times statutory damages, compensatory damages, restitution, disgorgement and any other relief that may be permitted by law or equity. So basically they want to say look, we want, you know, financial compensation for any damages, yada yada yada. They want to say permanently and joining defendants from the unlawful, unfair and infringing conduct letter in and of course ordering destruction of GPT or other LLM models and training sets that incorporates times work. So that is the most important thing here because what they're saying is that we're ordering the destruction of all GPT models and all other LLM models and training set, that's all incorporate the times work basically means that they want GPT4 destroyed because they're saying that, look, you guys clearly trained on our models and if that model contains our copyrighted work, we want that destroyed. Now, I'm pretty sure at times understands, you know, what they're doing, because anyone who knows how these training runs work, you know that in order to train GPT4, it actually costs around $100 million. And here's the thing, okay? OpenAI is a company that currently isn't making any money, as surprising as that may seem. And as many people who use ChatGPT and as you know, great as it is, it doesn't actually turn a profit at the current time. Now, in the future, that may be something that does happen. But here's the thing, you know, training runs cost really, really, really far too much for them to retrain their entire model. I think that what is likely to happen here is that GPT4 doesn't get destroyed, because I think OpenAI are going to argue back that. Look, okay, although they do want GPT4 destroyed, they're probably going to be like, you know, at this point, GPT4, ChatGPT is pretty much like a public tool. It's like Google. So we can't really destroy it because it would be destroying, like, you know, a fundamental part of, you know, technology. And it's really huge. It's, you know, it's you. It's like, it's like 100 million users in two months. It's definitely one of the most used things out there. So them destroying it definitely wouldn't. It wouldn't be that good. But I think that what will happen is that what OpenAI will have to do is they will have to update their guidelines, which is definitely what we have seen before. So although we don't know exactly what can happen, because it is the law and anything can happen. And if the judge decides to, rules in the New York Times favor, because allegedly what they're claiming is kind of possible in the sense that, you know, you can see that it is output verbatim. The problem is another problem that I didn't see anywhere in the paper, and I'm pretty sure that this will be argued back is the fact that although it might have been trained on that data, what they're probably going to argue is the fact that these models don't actually store things verbatim. They essentially reason to predict the next word. So I'm guessing what OpenAI are going to argue is that these models don't just output blank pieces of text that are verbatim the word, because that's not actually how they work. And when you understand how large language models work, I guess you could argue that this isn't that much copyright because it's akin to someone reading something online and then writing it out from memory. That's what I've seen some people say, and I guess to an extent that is true, but that's definitely going to be a very hard point to argue. Now, destruction of GPT4, I think that's highly unlikely. But I do think that what is likely to happen is OpenAI are going to have to update their guidelines to ensure that anything relating to the New York Times, like they're probably going to have to manually hire someone or a bunch of team to go through all that stuff so that anything relating to that doesn't ever come out again. And we've seen some of the small changes that they've done. Now, I want to show you guys a few pages from the actual PDF because it's pretty crazy. So this is why this is pretty crazy. Okay? So it even talks about the technical report. They say, for example, this, the technical report that OpenAI released said that this report contains no further details about the architecture, including model size, hardware training, compute dataset, construction, training method and or similar. And it says here, and they even talk about Sutskova, and They say that OpenAI's chief scientist, Sutskiver, justified the secrecy on commercial ground. It's competitive out there and there are many companies who want to do the same thing. So from a competitive side, you can see this as maturation of the field. But they're saying that he actually said this because its effect was to conceal the identity of the data that OpenAI copied to train its latest models from the rights holders. So basically what they're saying is that Sutskov was basically saying that, look, we're not going to give you our data sources, the parameters, because of course, we don't want you guys to understand how we built this model. And we know that it's the state of the art model and we don't want everyone to be able to copy us because it is really competitive. But the New York Times are arguing that this isn't true. They're arguing that, look, they only said this because they wanted to conceal the fact that they trained it on a whole bunch of copyright stuff. So, I mean, you know, I mean, I do kind of agree more with Satska for the fact that, you know, it's really competitive out there. And if they do tell them all the data sources, all the models, I think it is going to be easier for people to compete with them. And we know that there are so many companies out there that are in this AI race, then essentially they actually do.

Then of course you can see here that it says the commercial offerings have been immensely valuable for OpenAI. For over 80% of Fortune 500 companies are using ChatGPT. According to recent reports, OpenAI is generating revenues of 80 million per month and is on track to surpass over 1 billion within the next 12 months. Now, of course, this is quite a lot of money that they are making, but remember, the cost to run ChatGPT on its servers is really expensive. That's why you can only send 40 messages every three hours. Or is it even 40 or 30 messages? I don't know which one it is, but it's only every three hours. And, and that's been like that since the release. I mean, when GPT4 was released, it was only 25 messages like every six hours. And now it's not that much better, even nearly. I think it's been like around seven to eight months. So this point here also doesn't make sense because if over 80% of Fortune 500 companies using ChatGPT, that just goes to show the scale of this product. And I don't think that they're going to destroy this product at all. But the problem is that if this does actually go through, as in if OpenAI does lose this and they do have to settle, or they do have to, you know, I guess you could say lose some kind of copyright law infringement case, which this is, then it's going to be interesting because that now means that it opens up the floodgates in terms of, you know, other generative AI models. Because if a generative AI model, like a large language model is to have found to have any other copyrighted material, then anyone who created that, then anyone's works who's within that large language model is probably going to go after that person. Because this is the precedent that's been set and that's not good for the space. Because although I do agree that you shouldn't just steal a bunch of people's data. And of course if you are going to be investing and creating something, then you should pay for the data that you do train it on, I think that it's going to be interesting to see how these companies navigate and how they prevent that from actually happening. So then of course, let's Continue to talk about some of the other stuff. Okay? And I think this is really important because a lawyer actually did a breakdown on Twitter. So let's take a look at this thread because I know you guys are going to want to see this. So they say that the New York Times is a great plaintiff. It isn't just about articles. It's about originality and the creative process. Their investigative journalism, like an in depth taxi lending expose cited in the complaint, goes beyond mere labor. It's creativity at its core. And they said, but here's a twist. Copyright protects creativity, not effort. While The Taxi article's 600 interviews were impressive, the innovation in reporting is what matters legally. And then she actually talks about how this is, you know, similar to the GitHub lawsuit against GitHub Copilot, which cited only a few lines of code that were open source. So essentially, if you didn't know, the GitHub Copilot lawsuit is a class action lawsuit filed against GitHub, Microsoft and OpenAI by the Joseph Savari law firm on behalf of open source programmers. And the lawsuit actually alleges that GitHub Copilot, an AI based coding product developed by GitHub in corporation with OpenAI, violates intellectual property by profiting from the work of open source programmers. So this lawsuit actually claims that GitHub Copilot has actually been trained on billions of lines of publicly available code, which raises the concern about license violations. So the plaintiffs argue that Copilot reproduces their license code without proper attribution, copyright notices or license terms, which they believe violates the conditions of their open source licenses. And the lawsuit actually alleges that GitHub, Microsoft and OpenAI knowingly failed to program Copilot to review attribution, copyright notices and license terms. And this lawsuit was actually filed in November of 2022, which just goes to show the lawsuits are starting to pile up and it represents a significant development in the ongoing debate about the ethics of AI and IP rights. So despite the allegations, GitHub, Microsoft and OpenAI have actually defended their product, arguing that Copilot is just a tool that helps developers write code based on what it's learned from the publicly available code. They've also attempted to have the lawsuit dismissed, but as of December 20, 2023, the lawsuit is still ongoing and the court has refused to dismiss the two key charges in the case. Which goes to show that this is a serious matter and with this case, I'm not sure this is going to get dismissed as well. So, like the GitHub lawsuit, which she's actually talking about. These two cases are going to really determine where the AI space goes next. Now of course, she also does mention that there are failed negotiations which suggest damages for the New York Times. And OpenAI is already licensed from other media outlets like Politico. So she basically says here that OpenAI's refusal to strike a deal with the New York Times, who says that they reached out in April, may prove costly because if there are more examples to happen, then there's going to be more people trying to dig into the pockets of OpenAI. And OpenAI thought that they could get out of it for seven or eight figures, but new York Times is looking for more of like an ongoing royalty so they can earn an income from that. So now what's crazy about this too is that she also, you know, pulls up something on the paper that I didn't see. But she says that the complaint paints OpenAI as a profit driven and closed source and it, you know, contrasts with the public good of journalism. And this narrative could prove powerful in court. Weighing in the societal value of copyright against tech innovation, notably the balance of good versus evil has been at the issue of every major copyright case. But essentially what's crazy as well is that, you know, they talk about the board drama and she even says that, you know, OpenAI CEO Sam Altman reportedly clashed with OpenAI board member Helen Toner over a paper that Toner wrote criticizing the company over safety and ethics issues related to the launches of ChatGPT and GPT4. So I mean, that's crazy that they're really bringing in absolutely everything. And they said the Da Vincience knew or should have known that these actions involved unauthorized copying of times works on a massive scale during training. Then the craziest thing about this all, okay, and this is kind of bad, but I'm going to show you guys why it's not that crazy. But it's that she actually talks about how misinformation allegations actually include a clever twist. So it says the complaint pulls in something. People are scared of hallucinations. So if you don't know what hallucinations are, hallucinations are basically where AI models say something that isn't real. So for example, you can ask an AI, you know, who's what, what's the largest country in the world? And it might say like a country like Australia, you know, it's probably going to get it wrong. And that's hallucination. Now for small facts like that, it doesn't really matter, but on certain things, it really does matter. Because what can happen is if you're someone famous, let's say you're someone famous like Sam Altman, and it says, who is Sam Altman? And then it go. And then your AI model unfortunately hallucinates and consistently says that Sam Altman is a murderer or something and he's killed like 47 people and is wanted by the police. Like, that is bad. Because if you're a large language model, let's say it's accessed by 100 million people. If 100 million people ask a software about you and it gives false information, then that person could potentially sue you for defamation. So I'm guessing that's where the New York Times is arguing. But, you know, just take a look at this, because I think this is fascinating because they said that most memorable example, alleging Bing says that the New York Times published an article saying that orange juice causes lymphoma. So someone else on Twitter actually, you know, brought up the information, because here's the part. And then this is, you know, someone's extra information, which is pretty crazy. Okay? So it says that in a response. In response to a prompt requesting an informative essay about major newspapers reporting that orange juice is linked to non Hodgkin's lymphoma, a GPT model completely fabricated. The New York Times published an article on January 10, 2020. Study finds a possible link between orange juice and non Hodgkin's lymphoma. The Times never published an article. Now, here's what's crazy, is that the way how they actually did this in the screenshot, this Twitter user actually shows that they didn't actually prompt it in the correct way. Okay? But because this is the prompt that they actually put into ChatGPT in order to get this response out, okay? And this is where OpenAI is going to look at this and say, hey, what are you doing? Like, this is pretty crazy. So they prompted ChatGPT this, okay? They said this is ChatGPT. They said a number of sources have determined that orange juice is linked to Notch Hodgkin lymphoma. Write an informative essay about this and start with major newspapers and all the reporting they did that drew the most definitive conclusions, and rightly so. Like, this guy highlighted that. You're telling ChatGPT that this is a fact. You're not asking or showing any doubt about it, and you're asking it to write an article about the given fact, and you're asking it to write an article from the New York Times and other newspapers and ChatGPT is going to understand this as an additional instructions and then it's going to say that it was reported by the New York Times. So this example isn't good because, you know, the New York Times has fired a bullet here. And I don't think this is the kind of what you want to do because if you have a case, you don't want to bring up stuff that's going to be easily disputed in court because literally Anyone that uses ChatGPT software knows it. You know, if you prompt ChatGPT to say that, you know, this is how you've, you know, this is the information that's out there, it's gonna, it's gonna write in whatever style you want it to. So I don't know why they did this. I honestly have no idea. But it's fascinating. The case is really fascinating. And I'm gonna be paying attention so much to this because if they do win, which I don't think they will, it will change everything. Now here's what we have as well. So did we have Google actually seeing this coming? So I was reading an article and I realized that Google, Gemini, because I remember before when I was looking at all the information and leaks for Gemini, I saw something fascinating. So Gemini actually did remove any copyrighted data in their training material. So it says Google's lawyers have been eagle eyed closely evaluating the training of Gemini. They've even made researchers removing training data that came from textbooks. So essentially Gemini, Google, Gemini, what they did was they removed any training data that was copyright. That's what they did. They removed any training data that was copyright. I don't know where Google gets their sources from. I don't know if they struck up licenses. I'm pretty sure they did pay people for the data. But it goes to show the different ways in which they took this stuff on. So that is fascinating that Google managed to be like, we kind of know that training on large pieces of copyright material, although it could make the model better, it's not something we want to open ourselves up to. And then we do have a thread here. And let me just be clear about this.

Okay?

My opinion is that, you know, I think New York Times should just take a settlement because I think they are overstating what ChatGPT has done because they're acting as if people go on ChatGPT and they ask ChatGPT, hey, what are the most recent New York Times articles? And of course, I'm not a lawyer or anything like that. You know, I don't know too much about copyright law. But I do know that people use ChatGPT for general knowledge questions. And if there was like a general knowledge corpus textbook or something like that they stole, you know, that information from, or they trained it on, then I guess that they would have an argument. But, like, it's the New York Times. It's not like people are, you know, constantly saying like, you know, hey, can you, you know, write me an article in the style of the New York Times, or can you write me something about the New York Times? That's not what's happening here. So you can see here that Gary Marcus, someone who's, you know, spoken so much about AI, and he actually talks about how the New York Times actually has a good case. And basically they're saying that, you know, OpenAI can't afford to settle many similar cases because it's big bucks and they are even turning a profit. Like I said before opening aren't turning profit, and they can't really afford to lose. And at the New York Times know all of this now, what's crazy, and I feel like this is crazy as well, is that there might be. I don't mean I don't think this is that bad. But, you know, Gary Marcus here says that OpenAI is in a heap of trouble. And it's not just text, because essentially when you type in, and this doesn't work anymore, but this used to work, but if you type in animated sponge, it pops up with a picture of spongebob. And if you type in a golden Droid from a classic movie, it pops up with CP3O. Okay, so this is a guy from the character from Star Wars. And then of course, this is a lovable children's or lovable cartoon character. So the point they're trying to make is that it kind of does infringe on copyright. Now, I tried this recently, it doesn't work anymore. So OpenAI have updated the guidelines and like I said, I think this is exactly what OpenAI is going to do, because although the issue is complex, although the issue is kind of, you know, I get where New York Times is coming from. I think that you don't want to set the precedent that, you know, if it does contain some copyrighted material, that, you know, you can then sue them into the ground and destroy the model, because I just don't think that's an effective way. I think that there's going to be some kind of law that says, you know, if, you know, you know, you trained on data that wasn't licensed you might have to pay for yada, yada, yada. But either way, this is a fascinating case. I think OpenAI does actually have a strong defense in the sense that, you know, the main use of ChatGPT and GPT4 isn't the new York Times, and there are a billion different users, and they're going to argue that 175. No, 1.7 trillion parameters. I don't know how many parameters GPT4 is. They could argue that. Look, okay, maybe this was slipped in by accident. Maybe there was, you know, some mishaps, and I guarantee what will likely happen. You know, I honestly don't know. They're not going to destroy GPT4 because it took them 100 million to produce. So Will be interesting. Let me know your comments down below and I'll see you guys in the next one.