#15 - Katie Gardner, Partner Gunderson Dettmer

Speaker 1 Welcome to the Artificial Intelligence Podcast with your host, Dr. Tony Huang. Today I'm here with an IP Lawyer, Katie Gartner. Katie, can you do a quick intro about yourself and your background? Speaker 2 Yes, Thank you Tony for having me. I am a partner in the strategic partnering and licensing group at the law firm Gunderson Detmer. I've been practicing for 15 years but the last 12 of the years I've been here at Gunderson and. Benderson is a firm that operates primarily in the venture backed company space. We represent thousands of venture backed companies and the VC firms that invest in them. So naturally there's a huge focus on the technology sector in this area. And 90% of my practice involves counseling companies, beginning with founders at the seed stage all the way up to public company legal teams. And I advise on a range of issues relating to products and services, including the negotiation of licensing deals. Partnerships, strategic commercial deals in every space and I've spent a significant amount of time this year advising companies on AI strategy and risk as well as it and this is the other 10% of my practice advising investors on their investments into AI companies. So as legal advisors to the companies that are developing and the investors that are backing the most cutting edge products and and tools in the space, we are at the forefront of all of the. The new and novel legal challenges it's bringing. Speaker 1 So, given Microsoft's recent announcement on their Copilot copyright commitment for customers, where do you see the future of AI and copyright headed towards? Speaker 2 Yeah, so I think copyright is the most interesting and probably also the most uncertain area when it comes to. AI and what companies can do when they're training models. And maybe I'll just start with a quick overview of copyright. Sorry, I'm gonna pause for one second. Tony, is that OK? Or is this gonna mess up the podcast? Because I have to ask you a question. Yeah. Yeah, Sir, because I wanna know for your audience. Because in explaining all of the the cases, if you want me to walk through these cases, I was thinking that I do a bunch of background. On copyright in the fairy? Speaker 1 Oh, sure. Speaker 2 Doctrine. Is that helpful? Cuz I don't wanna go. Speaker 1 Yeah, so you can dictate the story. I just ask questions whenever I just wanted to. Speaker 2 I just wanted to make sure before I go to the weeds on this that I'm like, not. Speaker 1 I'm covering content that's like relevant for this is great. Yeah, feel free to to give the background on it. Speaker 2 Okay, okay. Yeah. So I think I'll just start up right now. So I think the most interesting or most uncertain area. When it comes to the legal issues around AI, model development is with respect to copyright, and copyright is a type of IP that protects original works of authorship as soon as the author fixes the work in a tangible medium of expressions. Photographs, books, movies, songs, computer programs and copyright gives the owners a number of exclusive rights, including the rights to. Copy the work to perform the work to display the work as well as what's relevant. For training a I models and using A models, the right to prepare derivative works based on the original works and copyright also gives the owner of the copyright the rights to authorize others to exercise as exclusive rights. Now to understand this in the context of a I, you need to understand how machine learning models are created. They're trained on. Massive amounts of data for language models or for images. Massive amounts of images and so on, depending on the particular medium of the model. And these models could be built and trained entirely unlicensed data sets, meaning that the company building the model has entered into an agreement with the company that owns the data, and they allow them to do that and to build a model. However, if that were the case, we probably would not have the very impressive foundational Llms that we've seen from Open AI and Meta and Google and and so on. And that's because on all those cases, those models were trained on what's called Common Crawl data. And that's just, it's the general Internet. So whatever data or content that you can find when you do a Google search is is possibly or likely part of the the training data set for these models. And so this means the companies are copying and using content without permission from the owner, potentially forming the basis for a claim of copyright infringement. And some people may ask, what's the difference between that and Google being able to show you content or images in their search results. And the answer to that brings us to the doctrine of Fair use. And that's ultimately the question that the courts are going to need to answer and and the numerous actions. That have been filed against Open AI and Alphabet and Beta and others in the last few months. Speaker 1 So I've always wondered this Why are major corps like Microsoft, Shutterstock, Adobe. They're introducing indemification clauses or is it called identification? Or is it indemnity clauses or financial? Either works great these clauses for the AI generated content, like. Why are they suddenly introducing these clauses? Speaker 2 Yeah. So that really comes down to the changes that they need to make to get enterprise companies to get comfortable using their tools. Speaker 3 Because. Speaker 2 There is so much uncertainty with all of the the, the cases that are currently being litigated, whether or not the. The models themselves in Fringe and then also whether or not the outputs in Fringe. And so by including these indemnities in their enterprise contracts, they're telling their customers, hey, you can be comfortable using codecs to write code because if you get an infringement claim, we're going to cover that for you and you can be comfortable using these these images generated. On Shutterstock, because we're gonna, we're gonna cover you if you get an IP claim. Now I think there's different there's different levels of risk when it comes to Shutterstock. I think they're not training models. I don't expect on the general Internet they're they're using licensed content. So it's it's much easier for them to get comfortable assuming that risk on on behalf of their customers and and is probably more of a of an optics type thing. But for other companies to do it, I don't think I've seen any of the other kind of large Llm's or Open AI or anyone agree this and I don't think they would because there's there's cases that are pending against them right now on these issues. Speaker 1 What What are some potential implications for like small business owners or startups that are leveraging AI without any protection in place? Speaker 2 It really depends on on how they're using it for certainly generating, generating content that you use in some public facing in some public facing way and depending on the prompts that you provide to the AI tools, you certainly could potentially create infringing content. Content, that is. Substantially similar to to copyrighted content, and you may be subject to a claim from from the author or the creator, I think for the vast majority of use cases, however. Speaker 3 The risk for for. Speaker 2 Companies that are leveraging AI is when their daytoday operations. Is not as much that I IP risk there. There are a number of other risks with using these tools including things that are inaccurate or biased or privacy and confidentiality risk. What are you putting into the model and how could it be? Is it protected? What are you, what proprietary content are you providing that may be then used by the tech company to to to train the model. So, yeah, so for the the small business, there's not the same IP risks that there are for the companies. That are building the models using copyrighted content and facing all of these cases. Speaker 1 In your opinion, like how do these copyright issues change the way companies draft up like their terms and conditions, especially concerning around AI? Speaker 2 Yeah. It has been interesting and it depends on what side you're on. So we've seen companies that have. Valuable content or content that they might want to, you know, monetize in some way have taken a couple different steps there. The New York Times got a lot of press when they changed their terms of service to say you can't use any of our content to train any models. Other companies have taken different steps by actually shutting down public. Access to their to their data. I think Reddit did this as those companies will will likely look for ways to monetize our content and enter into licensing deals and a couple of those deals have been announced. Open A I I think did a deal with shutter stock that was announced and a deal with Associated Press that was announced to be able to use that data for training purposes. So you have. Yeah. So on the content side, you have the company is trying to rein in and protect their data from being used without payment so that they can, so that they can retain the ability to monetize. And on the on the the model side, the companies that are providing these tools and where there's end users who are using them, you have the Microsofts and the shutter stocks who are taking one position where they are giving, they are indemnifying and agreeing to. To cover the legal risk but for the vast majority of especially for anything that's not enterprise just end user facing, I think the approach is mostly companies disclaiming any liability or risk because there are so many unknowns and things that they they just can't promise. Speaker 1 So I have, I've talked to a lot of companies where they've been leveraging AI for customer service such as like chat bots or voice AI. What are the legal requirements that these companies should be aware of, like moving forward? Speaker 2 Yeah, so if you're going to use, yeah. If you're gonna use an AI chat bot one, you have to look at the terms of the kind of third party tool you're using. Assuming you're leveraging a third party tool, Open AI, and you're not and you haven't built it internally. Because those terms may require you to make certain disclaimers or notices to your end users. There are also different laws, state laws, that may require you to make disclaimers that the the individual is not interacting with a human, they're interacting with with a chat bot, and then you should make sure you're doing things to. Mitigate the risk and a lot of this kind of depend depends on the particular use case in the way that you're doing it. But generally I advise companies to make sure there is an easy kind of process for someone to reach out to a human if something's not going right or they need help because they're it. It's well known that these. These chat bots are not going to get everything perfect and may may get things wrong and in addition to resulting in customer experience issues could result in some sort of financial harm or other harm if something they some output is is not appropriate. Speaker 1 Are there any legal challenges that are unique to voice AI as opposed to text based chat bots? Speaker 2 Yeah, I know. That's an interesting question. I think probably some of the voice or sound or even image stuff you have to think about, right? Of publicity issues and whether or not you're using it to impersonate someone else. And actually doesn't even need to be. There's right of publicity issues with respect to people who can. Monetize their persona. But there's also risks. Other types of risks aren't like criminal law. How are you using that voice? You're going to use someone else's voice to mislead someone or to commit some sort of crime. So yes, there there are different issues that would apply to that. Speaker 1 Yeah, I've seen these voice a I bots that are hitting like older citizens and trying to take their money through the phone. So obviously that's like super illegal, but now it's become a lot more automated. And I've noticed that people like older citizens are becoming prey to these boys AI bots. And I thought that was an interesting story that's been popping up on the news. Speaker 2 Yeah, it's pretty terrifying if you think about it. And I think there's obviously a lot of discussion happening within state legislatures within Congress on how to wrap. Wrap our arms around regulation. But separate from that, I think there there's going to need to be technical solutions to to some of this stuff where there are kind of these areas of real extreme harm because there's no way to know where how content is generated, what the sources were, whether or not someone is is using using content to impersonate someone. Speaker 1 So I'm going to ask you probably the most important question so far and feel free to dive a little bit deeper into it. So from a legal perspective, what's the best practice for integrating a I systems into existing workflows at a company without opening up any potential legal pitfalls? Speaker 2 I don't know if that can be answered in a single in a single question, so it's it all I guess. I'm a lawyer so this is like the classic. It depends on a lot of factors. What AI systems, how are you using them, what are the? Depending on what the use cases are will will will depend a lot of the risk around inputs and and outputs, but the key. It's the key things to to look at when you're and I assume we're talking about third party AI systems not AI systems you're building yourself are what what inputs are going are you going to put into the the AI system because you have to look at whether the data is proprietary or do you have the rights to do it. If it's your customer's data, what is your contact with your customer say? Speaker 4 Is it? Speaker 2 Personal data and look at your contract with those vendors and what are the rights. Speaker 3 Or. Speaker 2 Restrictions that vendor has with respect to the data hopefully they cannot use. I'll say it depends. It depends on the use case, but in most cases you may not want them to be training their own proprietary models on your data. In some cases you might because you might be taking advantage of. The features of that product by they aggregate and anonymize a bunch of industry data and that's the ultimate value that you're getting from it. So you might be okay with them training, but you have to understand what it is you're giving them and how they're using it. And then you have to look at what the outputs are and how those are being used in your in your business and make sure that your. Employees are very well trained on the. Speaker 3 Risks or. Speaker 2 Guidelines on on using the output and this is based on all the the very well known risks around a I and if it's something that's critical to get accurate. And the you're using a model that could produce inaccurate information, making sure employees know they need to review it for accuracy before relying on it, making sure the the data is not biased, et cetera. A lot of companies that we've been working with have been grappling with these issues and the the challenges with the fact that a lot of employees will just use this stuff because a lot of tools are free. Use this stuff without. Without any official kind of company authorization. So we've prioritized getting the high level talking points or high level bullets for them to share with our companies so that everyone knows what is and is not okay and then putting in processes for approving enterprise tools that are safe. Some companies and I I would say this is in the. Very small minority of companies that I've talked to have and I'd say even less so now. This was probably more March, April took the approach of, hey, we're just going to prohibit this altogether. I don't think that's a very practical approach because I think the reality is that's like prohibiting Google, like everyone's going to be using these tools and people need to know. Speaker 4 How? Speaker 2 To use them safely. And and to have them have. Speaker 3 Tools. Speaker 2 That the company's comfortable with, that they can leverage. There's a lot of benefits for companies who can figure out how to effectively and safely leverage these tools in the business. And then the other kind of best practice is looking at the enterprise versions of these tools, where you can do security reviews, you can get more legal protection, you can. Get better contractual terms around being assured that your content, your input contents not gonna be used to train proprietary models and the other any other things that your company may care about. Speaker 1 Are there any differences or can you spot any things that are unique between integrating internal AI products versus external AI integrations? Like, how does that How do internal and external AI integrations differ in their legal considerations? Speaker 2 Yeah, there, yes, there are differences with respect to external AI or AI that you're going to incorporate as a feature in a product or that a user of your product may be using. You're certainly opening up the door to more risk from of liability or claims. If if the AI gets something wrong, and especially if you're if you're leveraging, there's very few companies are building their own large language models because it's so incredibly expensive to do. And the large tech companies open AI, meta, Google they've already built these very very good products that that companies can use. So if you're using one of those models and you're fine tuning it, or using it for customers to provide their data, their data into it and not fine tuning it, making sure you have all the appropriate disclaimers. Some cases you may want them to directly accept the Llm's terms, or you may want to just surface all of the disclaimers so that you're not liable or responsible if the model produces something inaccurate or infringing or offensive. And then with respect to internal, I think it just goes back to the best practices of what I was saying before and looking at the use cases and how you're using it, what you're putting it. Speaker 1 Let's switch over to like recent news regarding Gen. AI. I read somewhere recently that amateur mushroom pickers have been urged to avoid like foraging books that were sold on Amazon that were written by AI chat bots. And they were saying that some of the chat bots were creating forging books about mushrooms and and I don't know if you know this but like mushrooms are very dangerous to pick if you don't know how to pick them correctly like you could die if you like pick the wrong mushroom and eat it. So what are like some implications from like a I generated books like that that are popping out in the wild and people are buying them and then they go out forging? And then they forge like the wrong mushroom, and then they die. Like what? Speaker 2 What happens, It's a interesting and very specific example. But yeah, that certainly that's certainly a very dangerous risk. I think I saw that line, a line about that somewhere, but to be honest, I didn't look too closely. I believe Amazon is requiring people to identify whether or not books or content has been AI generated. So I think this comes down to users. Users have to be somewhat responsible for knowing the sources of what they're relying on. I think the same way you wouldn't go on some some site that advises a bunch of around your health that seems sketchy or not right or isn't from a doctor and and kind of follow all of those steps. You should not be doing anything that could potentially cause death or serious injury unless you are very confident that the source of the information is valid and accurate and with anything AI generated it gets stuff wrong. That is a big risk when you're relying on outputs or using it ensuring contacts. Speaker 1 Awesome. Thanks for the background on that. And then also in recent news, I just pulled up a CNN report. It said a group of famous fiction writers including George RR Martin, John Gresham. They joined the Authors Guild in filing a class action suit against Open A I on Wednesday. So this was yesterday alleging the company's technology is legally using their copyrighted work. You wanna expand on that and give you give us your thoughts? Speaker 2 Yeah, this is there. I think there's been almost a dozen of these cases filed against Meta and Alphabet making these claims or similar copyright claims. So there was a bunch in June and July. Sarah Silverman, the comedian who wrote a book, was the name plaintiff and one of the earlier ones. So they're all they're all similar and I and I expect that we'll see more. Ironically. I think that it's the best precedent that we have for saying that training models may be fair use is from another Authors Guild case against Google from from 2015. I think in this case obviously the the authors are are hoping that it comes out a different way because then in that case it was found that the Google Books project was was fair use. In these cases, the authors are they're saying these Llm's have have been built by graping massive amounts of information and data that's public on the Internet, and made copies of their kind of copyrighted works in the process, and that they're unauthorized derivative. And on the other side you have the the tech companies that are saying copyright only protects expression not the underlying idea. Meaning that it it protects the way an author expresses an idea. But facts and ideas themselves are not protectable in the in the way and models work to use trained data to understand, generate and predict new content which is is similar to how the human brain works, only I can do it much faster and over much much much bigger sets of data. The the tech companies are also saying every output is not a derivative work. You need to have substantial similarity in order for work to be infringing, and most of the output that's that's being produced does not share any of that kind of copyright protected expressive elements with the the original works. And then last, you're saying even if this is infringement, it's clearly fair use because it's transformative. It does not supplant the the market for the original copyright works. People aren't going to use these models instead of buying these books. And that's what the Google Books case turned on that scanning those books for for allowing readers to search for author information and publisher information without creating kind of substitutes of the books. Speaker 3 Was fair use. Speaker 1 What about the the class action lawsuit against stable diffusion? Is that the same type of? Are they using the same type of argument? Speaker 2 In that case, yeah, no, I think those cases are actually a bit different and I have to make sure I have the the one right because I know there's a the Getty Images case. In that case, I think the images in in addition to like copyright infringement claims, I think the images actually reproduced the the Getty kind of watermark and you can see it in the image and there is a little bit more similarity between. I think this is one of the one of the challenges with image models versus text models. But the examples that they showed in those there, there's a lot more. There's a lot more clear overlap between the input and the output, and they also evidenced it by showing that the watermarking was was clearly visible. Speaker 1 Yeah, to my understanding the typically when you build. Data sets for these these large language models, you you them offshore. When I say offshore, we're in the states. So offshore in regards to the states just because it's cheaper. And what they'll do for images is usually they'll crop in a little bit so let the watermark will disappear. And I think what happened was like they just got lazy and they didn't watermark if they remove the watermarking with cropping which is what caused this problem. But I've seen the workflow of this before where an offshore team would go and manually start cropping images. So that's just a little bit of content. I'm also curious so there there's another like statement here where Open A I is pushing back on the class action lawsuit and they asked a San Francisco federal judge or court to narrow 2 separate lawsuits from authors including Silverman like alleging that the bulk of the claim should be dismissed. What What is the purpose of trying to narrow the the lawsuits downs It's just to remove? Speaker 2 Like these additional claims, Yeah, I guess so. I'm, I'm actually not a litigator, so I'm sure there's some litigation strategy behind this and I'm probably not as in tune to that would be my guess or trying to narrow some set of claims that they think are potentially easier to dismiss at this stage. Speaker 1 Gotcha. And moving on in the future, like how long do you think this class action lawsuits going to take and what do you think the outcome might look like? Speaker 2 These things can take a very long time. I don't want to say a decade, but it could be close to that. If you think about this, going through many cycles of appeals, whether or not the stuff ultimately made it to or makes it to the Supreme Court is another that would take a long time. I. Think decades for this, or it might not. I don't think decades, but I think some of these things can take 686789 years. Yeah, it can take a really long time when you're thinking of many cycles. It goes through this court and then it gets appealed and then it could take a long time. Speaker 1 I asked that question because a lot of the clients that I talked to there's skittish about using Gen. AI and they want to know what the courts will say before they make a move and it if it's if the this lawsuit is able to end. Relatively quickly there then they're able to make a decision whether or not they can move into Gen. AI. But if it's going to take a much longer time, then it's they might have to skip this product and do and use something else to generate value for their company. Speaker 2 Yeah, honestly, there's not going to be a clear outcome before. I think companies need to identify the risks and mitigate the risk, but they're not going to be able to eliminate the risks. And in some cases, and again it depends on the use case, there may be more risks with not adopting it and not taking some very calculated risks with using some of these tools than with kind of just not using them altogether. I think we will see more the the companies that are building the models taking more steps to try to make as I make their case look a little bit better. I just thought it actually just saw today that opening and I announced Dolly Three and this version is going to both allow users or authors or sorry artists to choose not to have their artwork use this training and also will reject prompt requests that say can you generate something in the style of some artist? And so that's an example of them taking some action to try to mitigate some of these IP risks. They also in July made that announcement about the GPT bot and how people how websites could block it so that publishers could make choices to not have their data be included in in their language models. And then there's all the licensing deals that they're entering into and other companies are going to enter into to maybe forego the risk of there being more plaintiffs bringing these actions. I think we'll have, we may have an answer in on all of the issues and all of the points in in the cases for a while, but I think we will see. In the products and practices to get comfortable with different specific use cases. Speaker 1 Cool. I'm told that you've spearheaded Gunderson's AI Webinar series. Is that right? Speaker 2 Yes. I and some of my partners. I say a number of my partners have been very, very deep in all of these issues. I actually just hosted a breakfast round table this morning for For fun GCS to talk about these issues. So we did the Webinar series. We've been producing lots of content and tools and advising many companies on these issues. Speaker 1 So from the series, can you share some of the main takeaways or even surprises? That you found when you're like talking to people, especially concerning around regulatory landscapes and like risk management. Speaker 2 I guess I'll say a lot of company, everyone's still trying to figure it out. So I think most companies I talk to, when I ask what are you doing or what's your process or do you have a policy, their answer is we're working on it. We're still trying to figure out what we should do and what we should do it and rightly. These are hard, complex issues. Some people have said, I just close my eyes and pretend I don't know because I don't want to deal with it or address the risk. But they know they're going to need to very soon. Some people say yes, we have policies, but we know no one's following them. And yeah, so I think there's a range of of risk tolerance and approaches that people are taking. I think there's also we're in the VC space and our our venture capital investors, it's a risky asset class. They're they invest in companies to get in in the early stages to have the potential for incredible upside and not every company is a winner and that's very well known and so I think a lot of the. Risks associated with these companies or these like IP litigation, it's acceptable risk for those clients because if it goes in the in the direction that's in their favor there's. The possibility for really significant gains. Speaker 1 So if you've had recent interviews and features in the Washington Post and Venture Beat where you've discussed the legal landscapes of a I, are there any recurring themes or like concerns that you've noticed? Speaker 2 I don't know that there's been I think every one of those seems to be on the following some new major news item whether it's a lawsuit or a change in terms of service. And so I wouldn't say that there's necessarily been a consistent theme with with all of them. I do think the most, the most. There are a lot there are a lot of issues that we could talk about when we're talking about AI and AI risks there's there's the regulation there's societal risks there's privacy risk security risks. There's bias there's but I think the naturally I'm an I an IP lawyer but I'm I've been looking at all these risks. But what I talk to people about the most is these issues around use of data and information, as well as like the transparency issues and the lack of transparency and many cases about. What is going into the models and how the outputs are being produced? Speaker 1 So the next couple questions are about like recent company commitments that were announced such as like Microsoft. Has stated that they will assume responsibilities for potential legal risk concerning copyright challenges like how significant is this move in the industry. Speaker 2 I think it is significant. So it's particularly for for venture backed companies it's a big. Risk to to not know whether or not your code for your product. Is OK is safe. Every big acquirer will do a what what we call the black Deck audit or source code audit. Black Deck is a big vendor in the space to identify any sort of any sort of risks and typically will make companies fully remediate anything they flag before they'll close the transaction. So this is an issue that comes up in in every single financing. We start talking to companies about it as soon as they're formed to make sure that they understand the risks and liability and have good kind of open source hygiene as well as making sure they have the the rights to everything that they're that they're creating or putting in their product and using an AI tool, which I also understand that engineers really love. It really does take away a lot of the more mundane aspects of of coding and doing that work makes it a lot more difficult for companies to to know whether or not they're in compliance. And I could see a real hesitancy to to wanna use those tools without having some sort of comfort that they will be covered from risk. So I think Microsoft making this move and other kind of companies making this move that we will cover you, You could use our products kind of risk free so to speak is important for enterprise adoption. And probably I'm talking about kind of the startup companies and non public companies, but probably even more significantly with with bigger public companies that that are maybe even more risk adverse. Speaker 1 So I use a lot of images in my personal and and professional life, like with Shutterstock offering financial protection against legal. Liabilities for AI images and there's an Adobe's identification clause for AI generate art. How's this gonna change the industry standard for AI content licensing? Speaker 2 Do you mean the like the companies that are competitors to shutter stock in Adobe? Or do you mean for the licenses that Shutterstock is getting 2 images from the original content owners? Speaker 1 Yeah. How's the how's the competition gonna change? Speaker 2 Yeah, I think it puts a lot more pressure for sure on the competition. One, I think the reason they can Shutterstock and Adobe could do this is because whatever rights they have from in in all of the images, in all of the content, they either own, own it outright, own the copyright or they have licenses that grant them the right to use the images to to build and train models and to produce this output. So I think there's one the question of do other companies have those rights? Can they do it? And if not, will they take the risk of of matching these commitments when there's a lot more? Potential liability for them. Speaker 1 So the next couple of questions are for people that really don't know the law that well. Speaker 2 So like how is liability typically allocated when it comes to IP infringements that are caused by AI that that's a question that we don't have the answer to to yet. So I guess there's the all of the cases that are pending right now are against the companies that are building the models, not against. I haven't seen or hasn't made the news any cases yet about someone being sued based on output they generated from a model. There's probably some stuff percolating but certainly an individual kind of end user of of any sort of model or AI tool could create content that's infringing knowingly or unknowingly. I think there's there's ways you can prompt the models to to have them generate something that is substantially similar and then you could use it in some way that is infringing. I think it's will the depending on how you're using it and who you are and will the kind of copyright owner care. Again that's like very fact specific, but there could be like theoretical infringements. All of the cases that are pending are against the the companies that are building and training the models. So it's it's less about the output. Although they they do say in all of these complaints that the the the output being generated and talk about and it clearly is like relying on the copyrighted content, but it's more about the the kind of training of the models themselves. Is infringing and it's the. Again, not a litigator, but from a litigation strategy. It makes sense to me that for these class actions, they go after the larger tech companies building the models before anyone would start pursuing individual claims. Speaker 1 Imagine if I had a company that created like this AI product. How do I make sure that it doesn't inadvertently infringe on existing IPS? Is there? Are there any? Steps I can take to reduce the risk? Or is it still like the Wild West right now? Speaker 2 What kind of AI product are you envisioning? Speaker 1 Ooh, that's a good question. Let's say it's a like a content generator. Speaker 2 So you didn't create it using AI you. Yeah, I think that just puts you in the seed of all these other Llms. So steps that you can do to reduce to reduce your risk. You could train it only on licensed content, the licenses that give you the rights to train it. And if you get that, then you're you're like shutter stock in Adobe and you can feel comfortable that that you're not gonna be the subject of an infringement. Speaker 1 Complaint. That makes sense. So like some of the coding like AI generators out there now are taking like codes that are licensed from like very liberal type of licenses like MIT or Apache licenses. So you're saying like take only take content that has a very liberal license and then train upon that, right? Speaker 2 Yeah, I guess it depends on exactly what it is. When you said content generator, I was thinking like image like their image generator, there are different entities that may own copyrights to to large databases of images and you can negotiate license with them. It's the same thing like how Open AI is entering into these license agreements with The Associated Press in Reddit. Like they're, I don't know exactly how they'll deploy that stuff, but let's say that they were to build a model that only was licensed on or only trained on licensed content. Then I think they could get comfortable that that they say we're not infringing. And yeah, certainly there's those are proprietary licenses, but there could be things that are, yeah, that are more freely, that are more freely licensed. You mentioned the open source. There's like a little bit of a nuance on that one because there's there's also, there's a risk of when you generate code that you won't include the attribution and the attribution information that you're required to include, even if it's even if it's AI generated. And that is the subject of the GitHub case. I think that was filed at the end of last year. So that's a slightly different example. But otherwise, when we're talking about traditionally licensed stuff, it's about getting license agreements in place. Speaker 1 Oh gotcha. Yeah, I so currently right now I code for fun, but yeah I just remember. Every time I had to go and get code I would just go on like stack overflow or something like that and copy paste code and then I find out like that code's copy pasted from a different website and that website copy pasted from a different website. So we didn't really know where the code came from initially. And this was like way before Jenny I was even around. But now I remember, I did remember reading like stack overflows. Web traffic did decrease quite a bit since now like people aren't using stack overflow for asking coding questions, they're using like. Some type of LLM to help them code which I thought was interesting. Little tippet yeah. So now I've been reading up on on IP law and there's a term I don't know, maybe you could shed some light on it. What is fair use? And like how do you use fair use when it comes to AI generate content? Speaker 2 Yeah, so fair use is it's a defense to copyright infringement. So it's not an affirmative right, But if you are found to have infringe someone's copyright, meaning you you infringe one of their exclusive rights to copy it or perform it or display it, or created a derivative work from it, then you are not liable for infringement. If that particular use case was fair, use it, and the purpose of it was to balance the rights of the creator with the kind of public interest of acknowledging that, like all art, owes something to to what came before it. And the purpose of copyright is to advance creativity. So we don't want to say that every single use of copyrighted work is infringement and. There are four factors that courts look up when they're deciding whether or not a particular use is is, is a. Speaker 3 Fair use. Speaker 2 And the the first factor is generally the most important one, and it's called. It's the purpose and character of the use, including the uses commercial or for nonprofit educational purposes. And probably unsurprisingly, nonprofit educational and non commercial uses are generally more likely to be fair. But commercial uses can. It can also be fair if they're sufficiently transformative. And transformative uses are generally uses that add something new and don't substitute for the original work. And this was what was found in the the Google Books case that they the that project did not substitute any kind of any users of Google going and buying a book because they could search through Google and find a couple snippets. And there's some other factors. So like the nature of the copyrighted work, the more creative or imaginative a work is then the the more likely it is to be protected. Whereas factual works are have kind of less protection or use of those is more likely to be fair use. And then the third factor is how much of the work was used. And so courts look at both the quantity and the quality of what was used. So it could be using a larger portion is less likely to be fair use. But sometimes there have been cases where a very small amount of a copyright work was used but it was found to be like the most important part of the work and therefore infringement. And the last factor is the effect on the market for by that use. To what extent does the unlicensed use harm, harm the copyright owners ability to monetize or sell their original work? Speaker 1 And so as a I continues to grow and evolve, what advice do you have for businesses looking to use a I without getting entangled in legal issues. Speaker 2 So maybe it would be helpful to, it's just so I go in the right direction of do you mean to try to stay within the kind of area of Fair use or do you mean cuz? And I'll just say, I'll say this, when it comes to relying on the fair use doctrine, it's generally the device that I've always given to clients when it comes to fair use is you can't rely on it to build a product. I know, I think AI is like a very special use case, right. This is like a very transformative technology and this is an area where companies are going to take risks and for good reason, big risk, big reward. But otherwise, like the vast amount of precedent, unfair use and the way we actually teach trainings on it is showing cases that have seemingly near identical facts and them coming out in completely different ways where one is fair use and and one is not found to not be for use. So it's it can be really difficult to predict what what the outcome will be and certainly not something that you can easily rely on. I think there's there's precedent that that points to it being fair use in the Google Bucks case and then there's actually even some more recent Supreme Court precedent that some people are pointing to is saying hey this AI training with copyright content may not be fair use because there's a market emerging where copyright owners may be able to license their content and and the way that this getting very technical. But the way that the Supreme Court interpreted the transformative use test in that case, that was a kind of Andy Warhol case potentially narrowed, narrowed that inquiry a little bit. Speaker 1 So, Katie, if I needed to get in touch with you, how would I do it? Speaker 2 If you need to get in touch with me, you could shoot me an e-mail at [email protected]. Speaker 1 And then finally, for the listeners who are just getting started with AI and their business, what resources or strategies would you recommend for staying ahead of the curb when it comes to legal considerations? Speaker 2 Yeah, we have great resources on our website gunder.com. We have a whole section where all the these client webinars that we've done, they're all they're public facing and available for consumption. So anyone can go in and and watch them and get our overviews. Speaker 1 Perfect. Thanks for being on the show. And until next time, stay curious.