Listen on Apple Podcasts, Spotify and YouTube.
Welcome to AccelPro IP Law, where we provide expert interviews and coaching to accelerate your professional development. Today we’re featuring a conversation with advisor, attorney and author Liz Rothman.
Artificial intelligence remains a hot topic with the rise of generative tools leading to new conversations regarding copyright. Precedent for generative AI and intellectual property law is increasingly being called into question.
In this episode, Rothman explores generative AI and frontier technologies and the impact and relationships with copyright and legal precedent.
Listen on Apple Podcasts, Spotify and YouTube.
Interview References:
2:16 | Artificial Inventor Project. (2022).
6:26 | Getty Images Copyright Case. Getty Images sues AI art generator Stable Diffusion in the US for copyright infringement. The Verge (2023).
Supplemental Materials:
TRANSCRIPT
I. GENERATIVE AI & IP
Neal Ungerleider, Host: Can you tell our listeners a little bit about you and your background?
Liz Rothman: I'm an attorney, and an advisor working in emerging technologies. So, fields like artificial intelligence, blockchain, AR and VR. And in areas of the law such as intellectual property, data privacy, and healthcare. All of these very disparate areas seem that they are far apart, but they all do intersect in these emerging technologies spaces.
I've had a private law practice for about a decade. And along with my private law practice, I'm an advisor for the Cantellus Group, which is a boutique AI governance advisory firm and an advisor for a nonprofit called the XR Safety Initiative that seeks to develop privacy and safety standards that are adopted by world governments, international companies, and stakeholders in the digital economy.
We're going to talk a little bit about copyright and AI, and my work with the Artificial Inventor Project, which is a series of legal pro bono test cases and is really looking at the challenges of integrating these new technologies into existing legal frameworks. Especially those that have developed far before we could have imagined computers generating novel and creative content. However, now computers are generating such content, and we must really respond in a rational way and look at how we can evolve our laws along with this rapidly evolving technology.
NU: What are some of these potential IP and copyright issues that may come up as a result of generative AI tools?
LR: I think we'll get into quite a few of them as we go through this session. But really, the main one is that it is in most countries around the world, right now, it's very difficult to protect content that is generated by AI - whether it is autonomously generated, or generated even by what many people might think of as humans using a tool. Some of that content is also not protectable right now. So that's really the main issue here.
And then there are other legal issues, broader issues, that come up of infringement, of the content that are going into these models, if that is copyrighted and protected content, and then being misappropriated in other ways. But right now I think, for this conversation, it will primarily be about the output coming out of these models, and if that is protectable.
NU: Can you tell our listeners a little bit more about your work with the Artificial Inventor project?
LR: Generative AI has become the new buzzword right now. For the purposes of copyright and AI, it's really helpful to have images. Or even direct experience with image generators like DALL-E or Midjourney. Or text generators like ChatGPT, or something like Jasper, or any of the platforms that are out right now.
About a year ago, there were about three of them. Now there's really been an explosion of this technology. And there are now hundreds and hundreds of AI startups in different places that you can encounter. Generative AI technologies that did not exist even six months ago, or definitely a year ago. So it's a very interesting space.
The AI generated output can, again, be generated autonomously. That is, without human intervention at all, or through the use of a prompt. So a string of words, a sentence, or increasingly longer text that’s written in natural English language. And then, it will tell a model what you desire it to output.
So, we're talking about copyright and these outputs of the models and the work that I've been doing with the Artificial Inventor Project. In 2018, there were cases filed to register patents in the UK and the EU for AI generated inventions. Those were then filed internationally, for international patent applications, in about 18 countries around the world. It's created a series of test cases around the world, trying to gain rights for AI generated output. Or to push the boundaries, essentially, of whether or not this output is protectable.
The project is run by [AccelPro Guest] Ryan Abbott. And it has really, over the last five years or so, become an increasingly bigger project; taking on more and more different facets of intellectual property protection for the different kinds of outputs. And overall, it's intended to promote dialogue about the social and economic and legal impacts of AI and generate stakeholder guidance on the protectability of AI generated output.
And as I mentioned before, this is super important. Because right now, in many countries around the world, this content and the inventions generated by AI are not protectable. So there's ongoing litigation in many jurisdictions, and it has even reached the UK Supreme Court. And then, there is a case in the US against the Copyright Office, for refusing to register an AI generated work over the last couple of years. And those cases very slowly work their way through the court system. But essentially, that project is really trying to push forward this conversation and test the boundaries. Hopefully, show how important it will be in the near future to protect this output.
NU: For our listeners who may not be familiar, can you tell us about some of the copyright and IP issues around training data for these large AIs?
LR: That’s a whole other can of worms. There's a number of lawsuits occurring right now. Including one, Getty Images suing Stable Diffusion, which is a large-scale model for copyright infringement.
When the models are trained, they're trained on billions of images that are usually paired to text data, as well. A lot of those images are coming from the internet. And the same with the large language models; they're largely trained on all of the internet. So, it's interesting for those of us that were maybe teenagers or older in the nineties, when internet culture really started to proliferate. A lot of the nineties and 2000s information is what all of these models are trained on, and the history books written before them. It's an interesting fact; that's where the training data comes from.
So, we have stock image libraries or Google Image Search, whatever images were on the internet, that's where largely these early models are trained. And so, a lot of artists' work was in that pool of training data. For Getty Images, Shutterstock, and companies like that, some of their databases were mined as well, to train these models. Even though when images are generated, it's not a collage, they're not putting together a collage of an image, but the system has learned over a period of time how to generate, based on what its training data was. And so, if you train on a large pool of Getty Images, the output of that Getty image might have a watermark of Getty Images on the bottom.
And that's exactly what happened in this case. And so, some of the outputs had Getty Images watermarks on them, showing that's how some of the training data was put in. You might have copyright questions there. If you're an artist and you input your name, you might be able to replicate something that looks like it could have been your work. Because the model was trained on the history of images and artwork. In those ways, there are a lot of copyright questions that come up. And, we're only just going to start to see the litigation coming out on that front. And where it will all end up, I don't know. The UK was going to allow for a broad exception for text and data mining on the internet - even for commercial purposes - that you could basically train on any data that you could get your hands on. And they have pulled back on that now. And they're going to slow down, and look at it a little bit more closely, before they make that determination.
NU: I remember you recently discussed in an article whether weights of large machine models are copyrightable in the US. Is this still an open question?
LR: Yes. I think that is going to be an open question for quite a while. So, a weight of a large language model is basically a combination of everything when you're training it. And you're giving the model feedback - whether that feedback is coming from human interaction, or if it's coming from the machine giving itself feedback. You create a collection of numbers and statistics about how to best generate images, text, content, through all the training process. So, those are what are referred to as these weights, and they're a secret sauce. They're the very special part of these models that will allow for the really high fidelity generation of images and generation of content that is something that you want to read. And that you want to look at, because that's kind of the aggregate total of the training. But that is AI generated output in most cases.
So, there are several ways that you could look at the copyrightability of weights, which is why I posed that question. It's a very interesting one. Because as we move forward, there are many steps to processes that we take. And along the way, software developers have been able to expand rights over what is copyrightable.
But when we're talking about this AI generated output issue - if it is entirely AI created, then we, for several reasons, can't copyright the output. And as we move forward, and more and more things become AI generated output, then that will create a chain of issues going forward. And those weights, being an interesting example of something that is so important in these models, and so, proprietary - so, perhaps there's other ways to protect it. But for the copyright issue itself - looking at the practicality of not allowing for protection of this kind of output, is a question that I think we should really be examining really closely.
—
II. PRECEDENTS FOR GENERATIVE AI IN LEGAL CONTEXTS
NU: And for the use of generator AI as a creative tool, what sort of technological precedents exist for that?
LR: Do you mean for copyright purposes?
NU: Correct. For copyright IP purposes.
LR: So going back to the camera, that's probably the first one where you could argue that the camera really took over the creative expression part of the necessity for copyright purposes. And so looking at the camera, and then moving forward to using some of the different tools that we use, including Photoshop manipulation tools.
Even computer software taking over some of the different things that human beings used to do. So the threshold for copyright ability is very low. But for example, alphabetizing a phone book does not qualify for copyright protection or the algorithm to do that, right? So if you say that, “Oh, my creative expression here is that I've alphabetized all of this.” That's not a copyrightable thing, because it diminishes the threshold for what is original for copyright purposes. But when you're looking at this computer generated output, you really can pass that bar pretty quickly. That it is novel. It's not a collage of factors, looking at something else and copying it.
There are a lot of these, when you look at the camera, even like, filters on top of cameras. And this evolution of technology as we've gone forward. I think that that's the primary historical place to look for. How copyright law has been flexible in the past, and has moved with technology. And sometimes it takes a little while, and it takes a couple of court cases. And it takes some sort of pushing to understand that if the camera came out, it doesn't mean the end of an art form.
It doesn't mean that all painters are now going to be out of work. Perhaps some of the portrait painters will be out of work. But then art got way more interesting after that, and there were evolutions that kind of came after it. And then, those sorts of arts were mediums that were protected. And looking to the history of technological evolution: copyright law, and our intellectual property laws, and our laws in general, have typically evolved to be inclusive of these technologies. Because if they don't, then humans have to find ways to work around that, right?
Humans are very smart. So, if the laws will not allow protection of these works if they are AI generated, then there will always be a person in that loop. Priority is a person in the loop, there. But there will always just be somebody in the way, adding another cog to the wheel, to make sure that it will fit within the laws and be copyrightable.
So, really looking at the history here and moving forward, from the 1800’s and on and even prior to that. And looking how we can evolve these laws, hopefully in a somewhat expedient way as well. Because use of this technology is not stopping. The ubiquity of it within our daily lives is only going to become more and more significant.
NU: You’ve written and talked about the distinction under copyright law between creative works, which are protected, and facts and ideas which are not protected. You've also addressed the current distinction under copyright law for works created by human authors, and those created by non-human authors, as we just discussed. Could you elaborate a little bit on copyright protection for AI generated output, when it comes to language models?
LR: It's a little bit more nebulous, right? So there's a lot of talk about images, and could we find the perfect secret to putting a watermark in there and figuring out how we'll prove the provenance of this image.
Or we could put it on a chain, or have the token on a chain and make sure that there is some way to prove where this came from. The output of language models is the same for copyright purposes; that the output generated would not be copyrightable. But the ability to tell if written output was generated by AI, I think, is a much more difficult question. Right? Perhaps we'll be able to figure out the watermarking, and to protect those images that are AI generated, and show where they came from. But if you have written contents, you can always strip out metadata. You can always - I think, we'll find out if somebody can crack this code - but typically, you can always find a way to copy and paste and strip that text.
And then, there have been so many programs that have come out. Especially, teachers are very interested in finding out how they can tell if a written work is AI generated. And also, the models get smarter by the day.
If you go into Chat GPT, you can prompt it to give you an essay on whatever a fourth grader might need to write an essay on; or a middle schooler and then you can say, “Can you write this in the tone of a fourth grader?” And you could even ask it to make some mistakes in there. You could really manipulate the output quite a bit from the text prompting that you're putting in, and really ask it to give you something that is difficult to tell if it was written by a computer, or not.
And I think, from what I've seen at least, telling whether something is AI generated or not, it's really about 50-50. It's a guess. And sometimes you can tell a little bit more if nothing has been changed. Perhaps there's some cadence and style that's really indicative of a certain model. But then, there are so many models. And increasingly, you could run these models on your MacBook. You could have your own model that you're tinkering and playing with. But that's also very difficult to, I think, show definitively. Especially if you're trying to prove that somebody's cheating, by using an AI software. It would be very difficult to prove that, in important circumstances.
You might be able to prove it if somebody is just, for a school project, putting it in and outputting it directly. And, oh, the cadence and style here is very much ChatGPT, perhaps. But at a higher level, I think it'd be very difficult to prove something like that.
—
III. FRONTIER TECHNOLOGIES AND THE LAW
NU: Let's talk about your career. You're working on some very interesting legal topics that intersect with emerging technology. Can you tell our listeners about your career path and what led you toward this unique place?
LR: I was a healthcare lawyer. I still am, for part of my practice. And, always interested in technology. And, maybe about five or six years ago, I got a little bit more interested in certain aspects of emerging technology, and some of the big changes that were happening. And, my path - which, I don't know if it would be the path for a lot of people - was that I dove right into the technology end. So, I took a computer science course, which had not been something that I had done before as a lawyer. And then, I taught myself some basic coding, and went that direction. And then I got really interested in different aspects of these technologies. And, especially the intersections of technology. Especially the intersection of AI, and ultimately blockchain, and augmented and virtual reality. And, I think that those are some areas that are really - when you can be at the center of those three - coming into the next decade, I think that they're going to be incredibly important.
And then there’s the various positions and consulting work that I do. When you get into these spaces and start talking to people, I think that there's not really a traditional path to follow. Although, I hear that there are lawyers going into law school right now, to go into this kind of work, which is very, very interesting to me. And, I think that there will be some more traditional avenues there. But I think, just diving in and getting involved in the communities around these topics and then, diving into whatever works, seems interesting. Because it is really charting a path forward into these - they call them frontier technologies, and that's what they are. Right? They're really on the edge. And so, you have to find your way on the edge. Which is a little bit different than some other paths, especially in the law.
NU: And does the healthcare law background help you with cases involving frontier technologies?
LR: From a patent standpoint - for sure, yes. But, I think a lot of lawyers and regulators and everybody think that these areas - because they are somewhat foreign to people - they're foreign to everybody in them as well, right? Because they're constantly evolving and changing, they don't fit within many of the boxes of our law right now. And, even from the copyright standpoint, I think that disclosure is the way out of this. Like, disclose everything, say it’s AI generated and allow it to stay within the system. Keep the system as it is, and move forward. And I think that there are certainly changes that need to be made, and there are realizations that need to happen, as a society, as we move forward. But also certainly in the regulatory space. So in the healthcare regulatory space, in the patent space, and then in contracts, a lot of this requires licensing and contracts to move forward.
And so, I think it is not as boring of an area as many people think. It's just technology, right? It's getting past the point of being afraid of any of the terms, and the technology. And once you get over that hurdle, then you start to see how similar some of the areas of law are. And that a lot of it is either just business law, or it's just the regulatory environment. Or it's intellectual property, and that the changes that are occurring are about understanding what's happening. And if you can keep on top of the speed that some of the changes are happening, and the new developments, then the legal part isn't as different as many people think.
NU: Do you have any advice for our listeners who are just getting started in their legal careers?
LR: I think it's a matter of figuring out which direction you want to go in. And I think, in this space in particular, it's about networking. And getting out and meeting people, and figuring out what's going on. Because - I think it's the case in a lot of areas - but if you really want to be in the emerging technology space, you have to really get out there and talk to people.
Right now, especially, there are hundreds and hundreds of new AI startups all the time. But try looking at them a little bit more critically, and seeing which areas are actually going to be moving forward. Or which of them are going to explode based on the most recent VC funding rounds.
I think the space, for me, was really important to know that I fully understood. And, I don't know if anybody fully understands the technology, but I understood it to a level where I could have conversations about it. And, I understood the technicals enough where I didn't feel overwhelmed, or afraid to get into a conversation with anybody about it. And I think that, for me, was important. Some people are a little bit more bold, and can go out and have more highly technical conversations when they don't fully understand what's going on. But for me, that was important. And then, having the solid legal background behind it, will help you navigate the direction that you want to go in.
Listen on Apple Podcasts, Spotify and YouTube.
This AccelPro audio transcript has been edited and organized for clarity. This interview was recorded on April 11, 2023.
AccelPro’s interviews and products accelerate your professional development. Our mission is to improve your day-to-day job performance and make your career goals achievable.
Please send your comments and career questions to questions@joinaccelpro.com. You can also call us at 614-642-2235.
If your colleagues in any sector of the IP law field might be interested, please let them know about AccelPro. As our community grows, it grows more useful for its members.
