Legal consequences of generated content

0:06Welcome to Practical AI. If you work in artificial intelligence, aspire to, or are curious how AI-related technologies are changing the world, this is the show for you. Thank you to our partners at Fastly for shipping all of our pods super fast to wherever you listen. Check them out at Fastly.com. And to our friends at Fly, deploy your app servers and database close to your users. No ops required. Learn more at fly.io.

0:43Welcome to another episode of Practical AI. This is Daniel Whitenack. I'm a data scientist and founder of PredictionGuard, and I'm joined as always by my co-host, Chris Benson, who is a tech strategist at Lockheed Martin. How are you doing, Chris? Doing well, doing well. I'm really excited about today because there's so many questions you and I have brought up in the show without the ability to answer. And I know we might get some answers today. Yes. And actually this one, not only will it be super practical and interesting, but it's also a tip from one of our listeners who suggested this guest.

1:21So we're really excited to have with us Damian Reel, who is a lawyer and technologist with experience in litigation and digital forensics and software development. So welcome, Damian. Thank you so much for having me. I'm thrilled to be here. Yeah, I feel very selfish this episode because I just have like a million sort of like legal implications, copyright questions related to like generative AI, large language models, all sorts of things. But before we get into some of those specifics, I know over the course of this show, we have commented on various things that have come about, like GDPR and then California data privacy stuff.

2:04And now we have like the EU AI Act and and all of this sort of regulation stuff. And then you've got other things on the other side on the litigation front where companies are, you know, getting sued for code generation based on maybe questionable training of models and other things. So maybe before we get in, from someone who is an expert in this area and thinking about it all the time, how do you view where we're at in relation to AI technology and regulation and kind of the legal side of things? How are those things catching up to one another or outpacing one another? And where are we at now as opposed to like maybe a year ago?

2:48What's changed? Sure. And maybe before I answer that, a brief, I litigated for about 20 years. So I was a litigator for about 20 years. I did tech litigation. So I'm coming to this from a perspective as a lawyer, but I've also been a coder since 1985. So I have the law plus tech background. So for your listeners' benefit, I'm not just a stuffed shirt that doesn't know what he's talking about. I can walk the tech walk and talk the legal talk, if you will. So really, as far as regulation, having litigated since 2002, I've seen ways that the EU and the United States have tried to regulate technology.

3:22And of course, they've had various degrees of failure, I would say, largely because the three of us know exactly how lots of technology works. But sadly, the congresspeople and the regulators do not. And so it's really the law is by nature slow and trying to get up to speed on a fast moving area such as AI is very difficult. So I would say that, you know, if past is prologue, I don't anticipate much good things coming out of regulation of AI in the near future. Some of these things are related to like this generative wave of AI where people are generating a lot of content with AI. I know that also you have a background, you mentioned your sort of coding background, but you also have a background with generative technologies, you know, maybe not like some with large language models and other things, but I know you have a very interesting story of some generative things that you did with music.

4:20Could you describe some of that? Yeah, absolutely. So both my current state with my job that pays me money, that is with Vlex, where I'm doing lots with large language models right now. We have a billion legal documents that we're running large language models across and doing embeddings to be able to do outputs. For example, a legal research memorandum and eventually be able to provide motions, briefs, pleadings, that sort of thing. So that's my job that pays me money. For the job that doesn't pay me money at all, which you referenced, is my All the Music project. This is a project that I started with my friend Noah Rubin, who one thing your listeners might find interesting is that around 2018, I did cybersecurity.

4:55The biggest thing I did was that Facebook hired me and my company to investigate Cambridge Analytica. So I spent a year of my life on Facebook's campus with Facebook's data scientists and my former FBI, CIA, NSA people that worked with me to figure out how bad guys use Facebook data. Did that for about 50 some weeks in a row. The stuff we would do on Monday would make the New York Times and the Times of London by Friday. So at the end of a 14-hour day on the Facebook campus, I retreated to my hotel with Noah Rubin, my friend. And we were in the lounge and over a beer, I said, do you know, Noah, how we can brute force passwords by going A, A, B, A, C?

5:29I said, what if we could do that with music where we would go do-do-do-do, do-do-do-ray, do-do-do-me, do-do-do-fa until we mathematically exhausted every melody that's ever been and every melody that ever can be? And so he said F, yeah, but he didn't say F, yeah. So within a few hours, he had a prototype where he cranked out 3 ,000 melodies. To date, we've now cranked out 471 billion melodies with a B. mathematically exhausting every melody that's ever been and ever can be. We've written all those to disk. Once they're written to disk, they're copyrighted automatically. So we've copyrighted 471 billion melodies.

6:01And then we placed everything in the public domain to be able to protect to use to all my melody lawsuit defendants. And so the idea is that before my talk in 2019, every defendant in one of those lawsuits has lost. After my talk, which has been seen 2 million times, every defendant has used largely my arguments and has won. And my arguments are largely that maybe if a machine cranks this thing out at 300 ,000 melodies per second, maybe we shouldn't give one person a monopoly, which is a copyright, a monopoly of life of the author plus 70 years for what the machine cranked out in a millisecond.

6:33So that's largely my All the Music project, which goes to, it is in a sense generative AI. It's a brute force generative AI, but it's generative AI. But the real question is, if the output of that is copyrightable, essentially if carpet-bombed the entirety of every melody that's ever been, and if I were a megalomaniac, I would sue everybody, right? But I'm not. I put them all in the public domain. But if machine-generated works are copyrightable, these are the bad things that can happen. I love that story. I just want to say that. It's pretty fun. And because it's been seen so many times now, 2 million, I've been able to meet with some good friends now.

7:08There's, for example, the former chief economist of Spotify. I'm now friends with. And the guy who was responsible for the first commercial MP3 to be downloaded, Jim Griffin, I'm friends with him. So anyway, so it's opened a lot of doors for me. That's awesome. And one thing that comes to my mind as you're talking about that is, because I'm also a musician, and I think a lot of people would think of, oh, these sort of melodies or chord progressions or however you want to frame it. There's a very human element to that that involves creativity. And I think this would maybe be extended to if we think about knowledge work more generally, whether that's like you being a lawyer and writing briefs, or us being programmers and writing code, or marketers being marketers and writing copy, right?

7:55Now these generative models can generate a lot of those things in a very compelling, and even I think people would perceive it as a very creative way, however you think about that creativity and coherence. So in your own work, maybe it's on the lawyer side or the coding side. How is that project and maybe some of the work you do day to day with large language models, how is that shaping how you think about this sort of knowledge work and the output of humans versus the output of models? I would say two aspects to this. And I think that the word creative is ambiguous, as all words are, most words are.

8:34And it is, you know, when I generated 471 billion melodies, I was creating those melodies, but it was by new means creative, right? So in that way, you know, those are just mathematically exhausting everything that's ever been. So that is a very simple version of what large language models do, right? Just saying, what is the statistically next sentence, right? And so this really gets to the heart of what we think human creativity is, right? Is it creative as in brute forcing or is it truly lightning in a bottle creativity that we want to protect with intellectual property laws and other things like that?

9:13I think what we're learning from my project and from the large language models is that maybe human creativity ain't as special as we think it is. As an example, on a Tuesday, a jury found that Katy Perry had violated a copyright of a melody that sounded like this. That was on a Tuesday. On my talk on a Saturday, I said that that particular melody shows up in my data set 8 ,128 times. So Katy Perry got dinged for$2.8 million over something that I had thousands of times just through brute force. So was that melody creative? No, it was just a brute force. After my talk was made public, the judge actually went back and reversed the jury verdict, saying that melody was so unoriginal as to be uncopyrightable.

10:00Essentially what I was arguing my TED talk. I think going to the heart of what is human creativity, it's good that we have large language models and projects like mine to be able to go to the heart of there are some things that we should protect and there are some things that are just unprotectable because they're unoriginal. You're putting an obstacle against what was otherwise maybe the weaponization of IP. Is that a fair way of putting that? 100%. We're using it as a shield, not a sword. Right. Which I like because it's having seen a lot of, for a non-attorney as having seen a lot of IP concerns out there in business, sometimes you're just kind of like, there's not much there.

10:33Absolutely. And so I did that with the copyright side. My friend, Mike Bomarito, who was one of the guys who you may have heard beat the bar exam. They used GPT-4 to beat 90 % of humans on the bar exam. One of those was Mike Bomarito. Mike approached me and said, I love what you did with copyrights. Wouldn't be great to do that for patents. And so right now we're doing, I was doing all the music project. We're now doing all the patents project. What that project is, is going to be taking all the patents that have ever been filed, taking each of the claims for each of those patents, putting those claims together in vector space and clustering them that way, and then generating every possible combination of all of those claims in all those existing patents.

11:13So if anyone in the future tries to recombine any existing claim into a new thing, they can point to our thing as prior art to be able to say, no, no, no, Bobberito, Real Cats, and maybe Rubin. They already did that in 2023. You can't do that again because they did that as prior art. As both a coder and a lawyer who I'm assuming some of these things you're generating are generated, but I'm also assuming in your day-to-day work and in your day-to-day coding, There are portions of what you're doing still that are not completely generated, or at least that you're editing heavily. How has this type of work influenced how you think about your own job moving into the future, working alongside these models or at least in an environment where these models exist?

12:04So I will say, and anyone who works with me will agree, that I am a coder, but I'm a crappy coder. I'm probably one of the worst coders you're ever going to meet. So a lot of my work with large language models is in the textual area rather than the code area. But in the textual area, I'll give you an anecdote that answers your question. I was reached out by the editor of a large legal magazine to say, hey, Damien, I want you to write the cover story on GPT. I said, how long do you want it to be? He said about 17 double-spaced pages. And I was like, man, I don't have time because my rule of thumb is one hour per double space page.

12:35So that's 17 hours that I just don't have time to do. But then I realized, oh, wait, the topic is GPT. So what I did is I created an outline of headings and subheadings. That's about two pages worth. And I said to GPT, for each of the bullet points, give me four sentences, essentially a paragraph for each of the bullet points. And it out. That was my one for the day. Crapped out the 15 pages worth. And then I spent the next three hours editing, moving, adding, working with the text, not accepting the 15 pages outright, but working with the text and regenerating. And then I got it out the door three hours later.

13:06And the editor is like, oh, this is perfect. I don't need any edits. Let's get it out the door. So that took a 17 hour project down to three hours. That's thing number one, is that this is not just accepting the machine output as is, but it's really us using the output as an assistant, much like Copilot on GitHub is using it as an assistant, right? These are essentially pair coding, if you will, co-authoring with the machine. I was doing a talk with the U.S. Copyright Office Assistant General Counsel, and he was talking about the regulations that they're putting out, saying that if machine-generated, therefore uncopyrightable, if human-generated, therefore copyrightable, and if machine-generated, you have to be able to disclose what aspects of the thing is machine-generated.

13:47So if you think about if I were to file a copyright registration with my article that I just drafted. What extent was that machine generated? And what extent was that human generated? Because I spent three hours adding, editing. If the machine generated in a sentence three of the words that were unmolested by me, and the other 20 words in the sentence were actually mine, do I have to disclose what three words were machine generated versus the ones that I edited? And that's with text. How would I do that with music, right? If I said to the machine, hey, generate a melody and generate a chord structure and generate lyrics.

14:21and then I spent from 1 a.m. to 3 a.m. rearranging all those things and then getting out the door, if the Copyright Office said what aspects of that was human-generated and what's machine-generated, I would honestly say, I have no friggin' idea because there's no track changes with my DAW that I make my music on, right? There's no track changes. I didn't track changes on my lyrics that I messed around. So I think this idea of trying to bifurcate what is machine-created and what is human-created is a fool's errand, and we're really going to have to reckon with that. Well, Damian, I have some very selfish questions that I've been pondering over in my own life as I've encountered them.

14:59And hopefully this won't seem like popcorn questions because I think it's very related to what you're talking about. But I think practical developers are hitting these snags as they're developing apps with this technology that are entering a sort of gray zone. So I'd love to get your thoughts on a few of these. One example that I can think of is, you know, a lot of people are building chat interfaces. It's very popular now to say, oh, build a chat interface over a website or build a chat interface over documents or build a chat interface over data, something like that. But people are doing this very frequently.

15:37It's very useful. My question is that chat interface or those messages are generated content, right? And that's what the user is seeing. But I see this huge gray area where let's say that I want to chat with Harry Potter, right? I take the book of Harry Potter and I put it in my vector database and someone asks a question of Harry Potter and I go retrieve the content. I'm just injecting that into a prompt, right? And I'm sending the prompt with, I guess, the book content into a model. The model's outputting some generated answer, and I'm sending that to the user. Now, I'm assuming, you know, I'm no lawyer, but I'm assuming I can't sell a new copy of Harry Potter unless I have certain rights and agreements in place.

16:26But what if I put this, you know, interface up on the internet and I start selling access to it? So I guess with that kind of very real world scenario, what sorts of elements do I need to consider there? And what's known to have a good answer? What's gray area? What's kind of being maybe being litigated right now? I'm going to answer your question, and it's going to be a fun walk. So take a walk with me. Okay, perfect. So this walk, it's going to begin with the Google Books project. So you might remember the Google Books ingested every book that ever existed, perhaps also including the Harry Potter book.

17:02A bunch of publishers said, hey, no fair, because you can't ingest all these things because these things are copyrighted. Every one of these books is copyrighted. They sued. And then the district court and the appellate court, the second court of appeals, said, yes, all those things are copyrighted. But Google's use of that is actually fair use. And this particular type of fair use is called transformative use. That the use that Google was using was transformative to what the original purpose of the book was. Purpose of a book is to read it, enjoy it, et cetera. Google's purpose was to index it.

17:32to be able to create a word index, to be able to then search all of the books, and to be able to provide the end user with a snippet, say maybe a page or two of that. So because it was not, you couldn't, I as a user, couldn't use Google Books to be able to essentially replicate the book process, but instead I'm using it to search, that is a transformative use, therefore fair use, that is not an infringement of copyright. So that was back in the day. Now think about how large language models work. So a large language model, if you have the input, it is ingesting, say, the entirety of Harry Potter.

18:02But really what it's doing is placing those in vector space, right? It's saying that these words are similar to those words in vector space. And once that happens, largely it jettisons the thing, right? In copyright law, there is the idea expression dichotomy. Ideas are uncopyrightable. So if I have the idea of a man in a black hat fighting a man with a white hat over a woman who is tied to a railroad track, Those are ideas that are uncopyrightable. You've seen lots of movies like that. But the expression of the idea, any particular movie that has that in there, that is copyrightable. So ideas, uncopyrightable.

18:36Expressions of the ideas are copyrightable. So if you apply that to what's happening when the large language model ingests all the books, it's essentially putting all the words into vector space. So it's saying, you know, this is a Bob Dylan-ism, or this is an Ernest Hemingway-ism, or this is a Harry Potter-ism. Each of those are ideas, not expressions of ideas. And so really, it's taking the expressions and effectively jettisoning those in favor of the ideas. So that's on the input side. And then on the output side, one can imagine that it's taking those ideas, Bob Dylanism, Ernest Hemingwayism, and then it's outputting them in a new expression.

19:14And if you believe the Copyright Office, machine-generated output is similarly uncopyrightable. So we're kind of faced with an idea that inputs ideas uncopyrightable, outputs the expressions of ideas created by machines similarly uncopyrightable. To your particular point, this has not been tested in court. So, you know, a judge who, by the way, may not know what he's talking about or she is talking about might rule against what I'm about to say right now. But at least I would make the really good argument that the ingestion of the thing is extracting the ideas from the book. And that is a transformative use because if you think about Google Books, they were printed three pages or so verbatim of these books.

19:54That's way more bulk than just, you know, think about the vector space. It's not reproducing any expression, right? It's merely taking the ideas. So if Google Books is permissible, almost certainly the large language model should also be. And that's really what is being argued right now in the cases that are happening with the GitHub copilot case happening in the West Coast. and then the stable diffusion case in Delaware and there's others like it where if I were the lawyers in that, I would be arguing exactly what I just argued right now. So to your specific use case, you are kind of interrogating this copyrighted work, but I would make the argument if I were representing you that this would be a transformative use.

20:31And just like you as a human would have read the book and you can, as a human, could provide output based on that book. In the same way, a machine should be able to read that book and provide output that is just taking the ideas of the thing, not necessarily the expressions of the ideas. So with that new expression coming out that you're describing and it being assuming that the Copyright Office view stands, it's not copyrightable, that's a massively different way of producing content from before till now and in the future. What does that mean for business and the world at large, considering that's a major, major change in how everything works?

21:10How do you see the future rolling out if that were to stand? First, I think it should stand because otherwise my All The Music project would essentially carpet bomb all the music and we would just have machines creating new expressions that would essentially make human expression obsolete. So number one, I think it should stand because if not, we are in a world of hurt with a copyright. Thing number two is you're right that never before in human history have we had a machine that creates new things. We've had the printing press where we take my ideas and then I can replicate it a whole bunch of times.

21:43We have the digital revolution in this, you know, 70s, 80s, 90s, 2000s, where now we can replicate human stuff. But never before have we had a way that the machine itself is making new expressions of ideas. And so as a result of that, some of the smartest people I know that are thinking about this is said that the web, as it stands, is probably large language models are going to stop right around November of 2022. Because anything after that, you're going to have a whole bunch of machine-generated content that is going to be essentially large language model-created things. If you know about the tech, and I assume your audience does, because it's statistically likely, it is smooth.

22:19Humans are jagged in the way that they write text. And that's what detect GPT and others is see the jaggedness of humanness. Machine generated content is smooth, not jagged. So that's what detect GPT says. Could you define that real quick, what jagged versus smooth means in this context? Sure. Yeah. So jagged means random. Smooth means statistically almost deterministic. So the idea is that we as humans say random things and we put things in a way that maybe hasn't been said before. Whereas machine, at least an LLM machine, is going to be able to say, you know, what's the most statistically likely word?

22:49and therefore that is smoother than our jagged randomness. The idea is that as the machines are creating the smooth, deterministic, statistically likely next word, that is essentially as new large language model ingest that smooth text, it's going to further smooth the corpus and we're going to miss all of the human created jaggedness that is going in there. So some of the smartest people I know are saying that maybe the web as it stood in November of 22 is maybe that's the last time we're going to have a lot of human created stuff that is truly jagged because here and out, we're just going to have machine created stuff that is smooth.

23:24And one last thing I would add is that one of the last bastions of human created jaggedness that we have is the courts. Because it turns out that, you know, people have talked about us being in a post truth era and post fact era. There is a person that's literally called a fact finder and his name is a judge. They spend years trying to find facts and then they write things called judicial opinions that have found facts that have been battled over years in the courts. So one of the last places where we can find this jagged, almost certain to be human written thing that is actually based in fact in our post-fact world, maybe is judicial opinions.

23:58And my employer, Vlex, has about a billion of those across the world. So maybe as we think about what are new corpuses that the large language models can train on that are truly jagged, that are full of factual things and not bullshit that's on the Internet, that is unvalidated. This is truly validated, human-created content that is high quality that might be a source to be able to ingest. Maybe this gets a little bit back to your article example where you interacted with the GPT output to write an article. And at a certain point, it kind of morphs into its own thing. What portion of it is machine-generated, what portion isn't?

24:34I know this is also happening just from seeing things like people are generating, for example, like adult coloring books using AI models and posting those in an almost automated way to Amazon. And then like someone can literally order a book. I think of other examples maybe where, hey, this book was written a long time ago. And so the wording is really difficult. What if I used a large language model to rephrase it in modern English and then I just post that and start selling it? So how long will this be debated in terms of like the copyright around this? And what should be on people's minds as they're creating this kind of content that they actually want to commercialize?