Welcome to Invisible Machines, a podcast produced in partnership with UX Magazine that explores AI agents and the ways conversational technologies are impacting our lives, our work, and the forces that drive us. I'm Josh Tyson, co-author of Age of Invisible Machines, a bestselling book all about the orchestration of AI agents. And I'm joined each week by Rob Wilson, CEO and co-founder of OneReach.ai. Rob has been at the helm of enterprise AI solutions for many, many years. And this podcast is where we investigate bold ideas from our book and beyond with other AI insiders.
and the most interesting people we can find. In this episode, we welcome back Ed Claris, managing partner at Claris Law and professor at Columbia Law School. Ed was formerly in-house counsel at ABC Disney and The New Yorker, as well as senior vice president for all of Condé Nast Publishing. Ed's here to help us explore part two of the United States Copyright Office's ongoing report on copyright and artificial intelligence, which focuses on copyrightability. He brings us some critical insights into what these updates to copyright law mean, as more and more people are now using Genitive AI as part of their creative process.
You can catch up on our previous conversation with Ed exploring part one of the report in season four of the podcast. This new report sets some useful guidelines for which aspects of these collaborative projects between people and AI tools are copyrightable, but it also raises new unanswered questions about the edges between the humans and the invisible machines in terms of output. Questions that Ed Claris is here to help us better understand. A quick reminder here to please support our show by supporting our sponsors. This episode is brought to you through partnership with OneReach.ai, delivering turnkey private architecture for building AI agents for teams inside of organizations of all sizes. OneReach.ai allows you to use any AI models with enterprise grade security and no vendor lock-in. Learn more about orchestrating AI agents for your team at OneReach.ai.
All right, well, Ed, it's great having you back on the show. You know, last time we spoke with you, the United States Copyright Office had, I think, issued the first part of their report on copyright and artificial intelligence. And now we have part two, copyrightability. So what's changed since we last talked? What do we need to know about what the Copyright Office is saying now in terms of how generative tools and other AI tools can be used for creative pursuits?
Hey, Josh, Robin, thanks for having me back. There's some more clarity. I think there's a lot of questions that emerged from the most recent report, but there's some more clarity too. I think the strongest piece of news coming out of this report is that it's a confirmation that prompts alone will not almost ever lead to authorship. Whether that means a lot of prompts initially, just like a very specific request of the chatbot to provide an output, even if there are hundreds of prompts, even if you get a result based on certain prompts and then go back multiple times to continue to refine the output.
you still, the copyright office has said, will probably not be considered an author with prompts alone. And the reason they've said that is that the output is still randomized. The output is still not something that you can ultimately control. It's going to bring up an output that it decides based on your prompts. You're not authoring it. You're not controlling the tool. You're requesting that it do something rather than telling it what to do. And so I think that's one of the big pieces of news. And a lot of people think that somehow if they work really hard or they put in a lot of effort or they're really good prompters, that that's somehow going to lead to authorship. And that's just probably not the case.
The copyright office didn't specifically say it will never add up to authorship, but it was very loathe to find a situation where prompting would be adequate. Another thing that they said was that, okay, prompting alone may not be authorship, but once the output is there in your hands, you can manipulate manually the output and create a copyrightable work. You can yourself change colors and design and you can add enough originality to that. public domain work, that output to make it copyrightable.
And so they did say that. There are a number of questions that emerged from that answer, though. One, for example, is, OK, let's say you added a lot of stuff on top of this public domain, unprotectable, machine-generated work. Is the entirety, therefore, once you've added your stuff copyrightable? Or is only the stuff that you added copyrightable? So take an illustration of a cat with a cigarette and glasses, drinking a glass of wine. Let's say you've added, you know, rose-colored glasses and you've changed the cigarette to a pipe and you've put a hat on the cat.
all manually. Well, maybe those elements that you added could be copyrightable or the assembly of all of those, but is the cat's face copyrightable at that point because it was generated by the machine or are other elements, the color or whatever else you may have left alone, part of that copyrightable work? Or do you have to disclaim all of the things that came out of the machine and only take credit for the things you added? It's almost like this mask of masking out all the AI stuff and saying, this is what I did. And this mask with a bunch of blank spaces in it, almost like those cutouts you stick your head in like the...
You know, the board where it has a hole and you stick your face in the hole, you're like, okay, everything in, you know, in the mask is yours and then everything else is it. That's so strange, but I get it. Yeah. That's interesting what you said about the, the fact that you're, the randomness of it. So the fact that it's, it's not deterministic by nature, that it's probabilistic by nature says you didn't, it wasn't designed by your.
even if you detailed the prompt out, you're still, in a sense, turning the AI into the actual artist. And therefore, you're just the person who's ordering the art, even though you paid for it. Yeah. That's exactly right. Another really interesting component was this question of, okay, so let's say the generative AI...
created the source image, that cat with the glasses and the cigarette. What happens if it happens the other way around, where you create an image of a woman's face with flowers growing out of her hair, but all just in line drawing, enough so that it's original enough that if you were to submit that line drawing of a woman's face with hair of roses, it would be copyrightable. It would be original enough. Now, you put that image into an AI generator, and you ask it to, through prompting, make the flowers roses that are... strong color red and make her face high definition with like nice makeup on with lipstick and mascara, make her hair blonde and her lashes long. And it comes up with this image that's on top of the source image and it says, okay, here you go. So is that resulting image copyrightable or is it only the underlying source image that's copyrightable? And the copyright office in a very strange way said, well, the stuff that you added through the generative AI is not copyrightable because it was machine generated.
But that raises a whole host of really hard questions around what goes on in the industry today. Because, for example... That's what I was going to say. Like, what if you took my art and did that? Like, do you own it now? Yeah. What if that was my picture and you're the artist, not an AI? Well, then that's this question of what is a derivative work? And yes, the output is derivative of this source material. So no question, you own... If someone else, a human being, takes your art and adds stuff to it, they're making an unauthorized derivative work, which you can stop.
Okay. So that's for sure. It's really a question of a machine. What the machine output is, is not copyrightable. It's a derivative of work, but it's not copyrightable, which is... Right. But what it raises is this problem that we've had, well, a reality that we've had for a really long time, which is that, let's say you're making the next animated film, right? The way it works, it used to work in the old days, was that you'd have artists sitting in a studio drawing every single frame with the moving, I don't know, coyote and everything had to be...
drawn. Now, that's not the case. The artists will come up with a source image. They look at what the coyote looks like and the hair and the color. And then the machine, for a really long time, animation machines, will then make the coyote run and do all the sort of things that the coyote might do. So are all of those animated films that we all know and love only protected on the source image, not on all else that's generated from it because the machine creates it? That seems kind of crazy.
And so I don't think that the Copyright Office really thought through what was really at stake here and what's been going on in Hollywood for a really long time. Well, it seems like it's being positioned as sort of like an ongoing investigation by the office. Yeah. So that obviously poses some problems, I guess, as well. And I think the report kind of makes it clear that they view this as also like a case-by-case thing. So there's some precedent. And then is that just to say that they're probably trying to figure it out as things start appearing, like as lawsuits start popping up?
I think, Josh, that's a great point. And I think that they have walked back as a result of saying it's a case-by-case analysis. They've walked back what their prior position was, which anything is machine generated is not protectable. We've walked that back. However, this case-by-case question kind of leads to an ambiguity and unknown for those of us out there in the world trying to protect and copyright works. And the difficulty with what you just said, which is...
while we're waiting for cases, you can't have a case until you get a, you can't bring a case until you have a copyright registration. So first, the Copyright Office has to register the work and then you get to sue for an infringement of that work. So if you can't get a registration, you can't sue on it unless you sue the Copyright Office for failure to register your work. And that becomes like a pretty heavy lift for anybody who's trying to, you know, not spend a lot of money. You know, I think it was like a couple of years ago now at a Colorado State Fair, there was an artist named Jason Allen who won the digital arts competition with an artwork that he had created using generative tools. I think initially there was a lot of pushback that, you know, thinking that he had just fed a prompt and this came out and he turned it in. But when they asked him more about his process, I think he said he'd spent 80 hours like prompting, pulling the image out of the generator, then manipulating it in like Photoshop or whatever he was using, and then putting that back in. And it was a lot of kind of moving it back and forth between tools. If you're an artist using generative tools,
Does it seem to reason that you would do yourself a favor to maybe document your process well, at least in these early phases? Definitely. I think documenting is a great way to have evidence that what you've done is actually your creation, that you had control out of the output rather than it being randomized. Because there are a lot of tools that will give you choices, but you'll select a choice and you'll manipulate it enough to make it your own. So I definitely agree. It's weird to think that Photoshop should add a future for just documenting what you're doing.
that's what i've been saying to other people in the industry like we are making this so hard and complicated and such a burden on what it means to be an author in this new ecosystem that people are going to be like not wanting to do it anymore yeah it's like undermining the whole point of it right to protect yeah And copyright traditionally is, you know, very easy to obtain. It's just a minimal originality. And people weren't stressed out about getting a copyright. They're worried about getting a patent, but a copyright's easy. It should be historically. And people haven't had to like disclose all the tools they use and all the ways in which, you know, Hollywood uses a million tools. The filters and cameras and CGI and audio enhancements and audio sounds and any number of things that we've been doing for decades. Yeah, I was just thinking about that.
We're getting to a place in the world where we're going to talk to our things. We're going to talk to our camera. We're going to talk to our microphone. And it's going to be robotic. And now I'm like, well, that really gets great, isn't it? Because you're telling the camera to film you and it's tracking you. So you're not telling it specifically where to pan and where not to pan.
So it's making assumptions. But yeah, at some point, that's so vague. It's just voice control. Like if you press the button, it would be clearer that it was you. But no, now it's a voice command and it gets blurry. It is going to get interesting. But like you said, I guess this is where things just go to court and stuff. Yeah, you wait many years and people don't know whether what they're creating is actually protectable. And there's a lot of lack of clarity in the marketplace right now. Somehow this AI generation has...
has become, has like kind of swallowed the entire question. In other words, we've been using technology to create content for a long time and it wasn't a problem. I understand that this is a kind of a new era and it does need to be addressed. It's just, I wonder if we're making it too complicated. Yeah. Yeah. It's getting easier and easier to make art. That's great. But then we're finding the edges of what is your art, which is.
Weird, because how does a court decide that? When we talked last time, we were talking about OpenAI and them consuming copyright material to create NLMs. Now we've got Chinese companies consuming copyright materials and then open sourcing it. How does this change the game? Or does it, it's one thing to try to go after a US company now. What does that look like? Or is that fairly straightforward?
No, I think that going after an international company is really hard. And you got to go onto their turf and sue them there or try to sue them here, but then you may not get jurisdiction on them. And there's a bunch of limitations, you know, serving them, process serving, absolutely going to be harder, whether they're in China or in Europe or in the UK or wherever. And I think that you're right. They have probably scraped the internet just like all the American companies have done.
And what does that mean for folks? So it's the same thing. If you use it here, then just like we talked last time, if you used it, then you're in the chain, right? You're culpable along with the provider. But I was thinking about, and maybe it just makes sense. Like if it turns out that Chinese company ingested the output of OpenAI and then open sources it, and now you're consuming the output of that.
Right. Does that change anything or is it just, it all traces back to you and the original author and all the middlemen don't matter? The middlemen matter. I mean, anybody who, let me see, if I'm tracking what you're saying correctly, a person uses DeepSeek in the United States to create a work that's infringing of somebody in the United States. So I was saying that OpenAI consumes copyrighted data. And so that's going to do its thing. And as we talked last time, but then when nobody knows this, but DeepSeek is now consuming.
So it's using the output of OpenAI to train its model. Right. Yeah. And that's happening across the entire ecosystem. Right. And then that's outputting into a model that's open source. And I'm like, now I'm using it. Exactly. The hallucinations multiply and the possibility of like all kinds of misinformation happens. Yeah. No question. I mean, if you're reusing mistaken information, for example, you're doubling down on the misinformation.
Right. But there's no copyright. That chain still stays. If it turns out that OpenAI infringed, then it goes all the way back to you, even though you eventually are using it. from DeepSeek. It doesn't matter that one handed to the next, to the next, to the next. Right. I mean, it goes back to the stuff we talked about around like inputs versus outputs. Right. If DeepSeek, their inputs of all the training data is infringing, it's always going to be infringing. It doesn't matter how much it was done after the fact. No matter how much regurgitating is happening. Right. Right. And if the outputs, people are responsible for the outputs and the AI companies might be contributing to the copyright infringement by allowing for infringement, but the user who's
getting the output out is going to be the one who's also responsible. Got it. Yeah. Got it. I have another question. I don't know if this fits into the copyright world. I think probably not as much, but we get a lot of requests for conversations with people that companies that are continuously nervous about AI and the hallucination concept, right? Like what if it says the wrong information?
We're seeing a lot of folks just talking about how companies need to start thinking about using AI agents in the same way they hire people versus thinking about it in the same way that they're procuring a piece of hardware or software. And I was sort of thinking about that. Like if you hire a person, how do you guarantee they don't say something that you don't want them to say? When you use an AI agent, it's got that same effect where it's not consistently going to say the wrong thing. So like a human, it might say it to two people, but it's not deterministic, so it's likely not going to say the wrong thing to everyone continuously. And I was wondering, is there a higher liability risk around an AI saying the wrong thing versus a person saying the wrong thing?
They will. Yeah. So I think that starting with your base question, which is, do we need to be paying more attention? I think you're saying to attribution, authority, provenance of information. And should there be more standards bodies that are watching out for that? Should there be more metadata that gives information around provenance? All those efforts are being undertaken by the standards bodies that are typical at W3Cs of the world, I believe. NISO is another standards organization.
And they're all trying to get to that question of authenticity, provenance, accuracy, and accreditation. And they're also very aware that there's one thing if the AI is reading a CAT scan and telling you whether you have a clogged artery, and another if it's reading a birthday list and telling you what your birthday is. And it really matters. Is there more liability for the CAT scan versus the birthday mistake? It's like real damage that would be done. And so the doctor who relies on artificial intelligence to read a CAT scan is going to be him or herself liable for the result. It's not, they're not going to be able to lose the machine you told me and it's not my problem. Got it. Got it. So it's this idea that at some point there's a person.
that has to answer for something whereas with an ai it goes to the person who bought the ai it's it's and it traces itself back at some point to a human decision um i think so yeah okay yeah yeah and that makes sense i just um i just realized like there's as these folks are saying well how do we know they're not gonna you know hallucinate and say the wrong thing in some conversation and you know my reaction how do you know your call center agent isn't going to say the wrong thing in the wrong conversation. What's the answer? The answer is there's some level of risk that you have to take to be in business. And you have to decide what that is. You could just not talk to customers. That would probably solve it, but it creates other problems. And I just wondered if there was some reason that...
that there would be more liability or less liability with the AI agent making that mistake versus a human or if it's pretty much the same thing. Our customers are coming to us and saying like, the legal departments are like, well, how do we know that it's going to say what we want it to say? And we're trying to say, well, it's not deterministic. there's ways to narrow it down. So there's, you can use an AI agent to choose like almost like multiple choice. So you can pre-write the answers and then ask the AI to choose from a list of answers. So, but that doesn't work for all use cases, but that's, you know, there are techniques you can use to kind of limit it, but the more you strap it down, the less valuable it becomes. And at some point you're going, so then what? You bounce it to a person who now has the same liability. So from the, from the legal departments,
perspective, I was trying to think, is that an accurate thing for me to say? You're going to have the liability with the person or the system. And if the system's error rate is better than the human, it's kind of like driving, right? If it errors less often, it feels like even though it errors, it's less liability than a human that errors more. But I don't know if that's true. It sounds logical, but I don't know if that's true. No, I don't think it's true. No, the liability levels are not different. And I think you may be right. Machines will make fewer mistakes in the end of the day.
Yeah. Well, speaking of like, I guess, software creation in a way, yesterday I had the chance to watch Rob. He was using a platform called Cursor to create an app. And he was doing it all with, it was all very conversational. It was a lot of creating prompts and then switching, like kind of choosing between different LLMs to see which one would give him usable code that he wanted. Rob, maybe you can describe a bit of what that work looks like. But I guess the question in there too is like, are there any sort of copyright considerations that need to be made where software creation is concerned? Or is that, you know, is that...
Same thing, right? We talked about that last time a bit, I think. Software is a literary work. You're writing sentences and paragraphs according to the Copyright Act. And if you have a machine generating stuff that is randomized, then it's a problem. If you have a tool helping you with something that you can control the output of, it's not a problem. Yeah, makes sense. I think I know the answer to this, but I'm going to ask it for everybody's sake.
we're getting to the point now where you're prompt chaining. So let's say operator, right? Where you're like asking it to do things and click on that and go over there. And then you're capturing your transcript. So the series of prompting that you did word for word and then saving it and then reusing that to trigger an AI to follow the same prompting, right? So exactly. My assumption is that
the script you created is copyrightable, but the output's not. So if you created a, you wrote the script and it's, and you feed that exact script to them on that prompt is protected, but the output's not. Is that accurate? That's exactly right. That's, that's that concept of source material that I was talking about where the source material is protectable, but what, what the AI adds to it is not.
That's, yeah, that makes a ton of sense. Well, this is great. Thanks for the update. Sure, it was really good to see you. Thanks for having me again. All right, well, thanks for hanging out with us again here on Invisible Machines. Don't forget to like and subscribe wherever you get your podcasts. Always remember that you can watch this podcast too on our Invisible Machines YouTube channel. Head over to uxmag.com. We have a podcast page there where you can explore our back catalog and you can look at episodes centered around popular topics. Many, many thanks to our incredible production team here in the U.S. and in Ukraine. Elias Parker, Michael Lipinov, Allison Harshberger, Aless Lesarenko.
Virup Akhodkov, and Darina Moskovchuk. Till next week, Oaki.