This Week in Startups - Who Owns the Output? AI Copyright & IP Explained w/ Chris Paniewski | Startup Legal Basics

Episode Date: September 18, 2025

Today’s show:Jason sits down with Wilson Sonsini partner Chris Paniewski for part two of our Startup Legal Basics series on AI law — this time tackling the question every founder is asking: who owns the output of AI systems?Chris has advised on some of the biggest AI deals in history, and in this episode he helps break down how intellectual property law is colliding with generative AI.Jason and Chris cover:What kinds of IP rights might apply to AI-generated output (copyrights, patents, trade secrets)Why human authorship is central to copyright lawWhether prompts are protectable IPHow GenAI platforms’ terms of service affect ownershipReal-world examples from Star Wars to Wirecutter to WarholWhat startups should document when using AI in developmentHow investors diligence AI companies on IP strategyIf you’re a founder, investor, or operator trying to understand how to responsibly use AI without losing your IP, this conversation is essential.*Timestamps:(0:00) Jason introduces part two of Startup Legal Basics with Chris Paniewski(1:20) Star Wars example: using LLMs to generate existing IP(3:10) Human authorship requirement in copyright law(5:25) Prompts, super-prompts, and the role of creative control(7:45) Photography analogy & the “monkey selfie” case(10:00) Patents, inventorship, and AI-generated inventions(12:05) Startups balancing speed to market vs. protectable IP(14:15) How to layer human creativity on top of AI output(16:20) Fair use, derivative works, and investor diligence*Check Out Wilson Sonsini: https://www.wsgr.comCheck out all of the Startup Basics episodes here: https://thisweekinstartups.com/basics*Follow Chris:LinkedIn: https://www.linkedin.com/in/christopher-paniewski-09331a59/*Follow Jason:X: https://twitter.com/JasonLinkedIn: https://www.linkedin.com/in/jasoncalacanis*Follow TWiST:Twitter: https://twitter.com/TWiStartupsYouTube: https://www.youtube.com/thisweekinInstagram: https://www.instagram.com/thisweekinstartupsTikTok: https://www.tiktok.com/@thisweekinstartupsSubstack: https://twistartups.substack.com

All right, everybody, welcome back to StartUp Basics. Why did I do this series? I get asked the same questions over and over again. And I like to help founders navigate complex issues, even the simple ones. Today, we're going to talk about a complex and a moving target with Chris Pyniewski, partner at Wilson Sincini. You know, we talked, Chris, in my first discussion with you about input and training data. Great. If you guys want to see that, go to this week in startups.com.

slash basic. Part two, output. You know, and I brought up the very evocative issue of Star Wars content. One thing you might want to do using a large language model out there is say, hey, make me into a Jedi Knight. Now, is that legal or not? It turns out a lot of the LLMs have said, when you try to take IP from the Disney Corporation, which is one of the preeminent rights holders in the world, who is also one of the preeminent defenders of their rights, some of the LLMs now, Chris, I don't know if you noticed, will not let you do that. Right. They are protecting themselves, right?

They don't want to show that they can generate infringing content, and they are looking to be able to, frankly, buster their own fair use argument, which we talked about in the prior podcast. And if they put these kinds of guardrails in place, those are some of the best practices to try to protect the LLMs from lawsuits from third parties that they're not doing enough

to protect copyrighted content. And a key part of fair use and a key part of copyright law is, hey, if I made Star Wars, I'm George Lucas, 1977, and we're sitting here now, you know, 50 years later, you know, trying to mitigate

how does Star Wars work in the age of AI? It turns out, well, we had an analogy to this. DVDs came out at a certain point. VOD, video on demand came out. merchandising came out. All kinds of new technologies came out and different opportunities for George Lucas to, you know, generate revenue from his IP. And it turns out the language models are no different. AI will be no different. Just because the new technology does not mean you get the right

to take George Lucas's or now the Disney Corporation's opportunity, correct, Chris? Absolutely. And just building on top of that, if George Lucas had used AI to create the characters, to create what a Jedi looks like, it's questionable whether he'd be able to be able to, to protect it. And I think that's a lot of the things that we're facing now is when clients are using AI and they're using it to create, hoe, to create other IP they want to protect, does that deprive them of the ability to actually enforce their rights in it against others? Oh my God. I didn't even think about that. So if I said, make me, you know, a weapon that would be used in space in the science fiction future. And some LLM said, you know what? Storage exists.

let's make them laser swords. And I created the idea of a lights and you said, hey, give me a name for it. And then I try to IP protect that. This is a totally unknown area. Is it not? Well, so we're definitely budding up against what the law says that human authorship, let's say for copyright,

is what's required. That spark, that creativity has to have originated with a human. And because LLMs are doing a lot of the selection. They're improving upon your prompt. And so they're filling in a lot of the details. So a lot of times the creative elements are left to the AI. And unless the human has the ability to control that and actually influence all of

the creative elements, there's not enough human authorship there to actually make an output, let's say, of an image from an LLM copyrightable. And folks have tried to register it at the copyright office and it's been rejected. Interesting. So let's keep pulling the string here. And it's just so great that you're on the program to talk about this because, again, sitting here in September of 2025, we recorded this, we are literally in inning number one of this issue legally.

Now, if I said, hey, I have an idea for a weapon from the future. It's a lightsaber. It is six feet long. It has a hilt. It is powered by a khyber crystal. And it's red if it's a Sith and it's blue if it's a, certain type of Jedi and it's green if it's a good Jedi. And here's what it looks like.

Make me a sketch of one. Now I've given it a super prompt, and it's doing a little bit of a drawing that I could have given any artist to do. That might make its way through the description, at least, the copyright office. Am I correct, or am I trending towards being correct? So I think the more detail that you provide the LLM, the better.

I think the best analogy that we have is, some point, it was questionable whether photographs taken with a camera were copyrightable. And with the Supreme Court... Really? Is that true? Yeah. Oh, tell me everything. This is fascinating. I'm learning something here. And it went all the way to the Supreme Court. And what the Supreme Court

said is that because a human is involved in setting the scene, posing the individuals, playing with the settings on the camera and ultimately deciding when to push the button to take the picture, that is enough involvement where the camera is a tool and the

human is the creator. So it's all about retaining those creative elements at the human level and the human having the ability to influence them that is the hallmark of copyright. Right, because it would seem silly to say the paintbrush is the author of the Mona Lisa. Right. Just a piece of wood with some horsehair on it. Now you get to the camera and you said, well, if my security camera took a picture and I wasn't involved in it and then I took a screen grab of and I tried to submit it, I don't get that copyright, do I? Correct. And there's actually a case where somebody tried to register a picture taken by a monkey.

Oh, I've heard of this. Tell us about it. Yeah. So somebody tied to register a photograph where a monkey pushed the button and the shutter took an action. And that was not registrable. It was not a human operating the camera. Right. So there's a, there's a human requirement to copyright, and even for, patents. So somebody tried to register a file a patent with an AI as the named inventor. Couldn't do it. No bueno. Interesting. I took that personally. You know, I've been called a monkey by many people, but I thought I should have gotten the rights. I thought, you know, I had heard that story before. So I think I'm starting to build up a mental model here. If it's somebody else's IP and I do

output, no one no without permission. If I had permission, hey, that's an opportunity. And this is what I always try to talk to my founders about is if there is an opportunity to build that and the original IP owner isn't building it, well, here's an idea. Build a proof of concept, build a deck, and go to them and say, I have a new tool, I have an idea. I was wondering if you might be interested in our company enabling your company to do this, either through a license or on your website. So if Jedi Me was a really good idea, well, maybe that could exist in the Disney Plus app and they don't have time to build it so you could build it for them and create the characters and the artwork and say,

hey, here, it's done. Can we do a revenue share? Can I sell it to you as a SaaS product? Yeah. Exactly. So there could be partnerships that evolve, and AI can fuel that. AI can help.

But I think what also startups are struggling with is, well, everyone else is doing it. And if I don't use AI for development and I don't be the first to market, I'm out of business. So how do you effectively balance using AI for development, but having protectable IP? And that's what we often.

oftentimes coached clients through is, well, where is the value in your business? Is it being first to market and having great customers and a great following? And you're putting sort of less emphasis in the value of your IP? Great. It all comes down to, and this is what investors will ask, is what is your thoughtful approach to using AI in the enterprise? And what does that look like? So if you're using it in development, what steps are you taking to document?

What was AI generated? What was not? And then at what point are humans brought in the loop to try to actually make it protectable? Because usually the output is not usable as is. Humans are going to have to integrate it. Humans are going to have to make it work with other applications. And when you start doing that, then you actually start creating copyrightable subject matter. And there's actually an example from the copyright office where somebody took a comic book that was AI generated images, but they added text and they actually organized the image in a story.

So while the images weren't protectable, the assembly of those images as a compilation is protectable along with the text describing what the images are. So there's ways to layer on top copyrightable protections on originally non-copyrightable AI generated works. And this happened in the art world, if I remember correctly, Andy Warhol, Baskillat, you know, they were doing very interesting collage and transformative work. and they too were able to navigate that minefield, yeah? Yeah, that's right.

And I think one thing that those folks got tripped up on from time to time is the work they built on top of, so the work their work was derivative of, did they need rights to that thing? And so thinking about when you're building on top of somebody else's work, just like if you go and ask AI to generate something for you, that could potentially be pulling on someone else's IP, making sure you have all the rights necessary to do what you're what you're desiring to do.

Yeah, I remember he did these Campbell soup cans. And that was a really interesting case. The one I'm also, I've struggled with myself, is I'm a big fan of wire cutter. I don't have you ever heard of that from the New York Times. And I used it as my Bible. You know, if they say this is the consensus choice, that's my little hack, right? Okay, I'll just go with it. Sure.

And I was a subscriber to New York Times to do that. And then I was like, yeah, you know, I really don't go to the New York Times. time for too much. And people will summarize it, re-blog it all the time. I only need the headline, but wire cutter I kind of need. And, you know, one of my LLM subscribers, I'm not going to call anybody out here, would just give me, in a much quicker fashion, I would say, tell me what the consensus choices are across different review sites for my next, you know, phone, my next laptop, my next TV. And it would be like, wirecut it says this and consumer report says this. And I did not need to subscribe to those two services anymore.

This would be a heck of a case, yeah? Yeah, so I think that gets back to something we touched on earlier in the prior podcast of, is that fair use, right? So when the LLM is summarizing this kind of third-party content, are you depriving wire cutter, are you depriving wire of traffic and their market that they're trying to address? The answer is yes. Now, yeah, I mean, clearly it's an explanation point because I, I,

I would be like a witness or an actual example of that occurring. And when you do these cases, you do have to show damages, is my understanding, right? There has to be a party who has been damaged. And so you'd have to find customers to actually win the case. Am I correct, generally speaking? Yeah. So there's a combination of damages available. So either there could be legally prescribed damages, which are called statutory damages,

or there could be actual damages. How much did you actually suffer? And that'll depend on what is going to be more advantageous to the rights holder. So, yes, there are damages available. Now, let me throw you a curveball here, if I may, Chris. Sure. The LLM got that information

because somebody else made a summary page, a human, who said, I am debating consumer reports selection of the Android phone, you know, the latest Android phone from Samsung versus the, latest iPhone, and I've reviewed consumer reports and wirecutter, and here is my meta-analysis of why I think they were right and wrong, so that's a transformative work done by a human, and then the LLM indexed that page, and then gave me the information, and never took it off

the New York Times or Consumer Reports pages. Is the LLM in the clear now, if this could all be proven? Well, I think at the very least, the LLM probably was trained on the content from that personal review site. and that author who wrote that personal review has copyrights and network just like the New York Times had rights in the original review. So it's just, right, it's just sort of layering. Yeah. What if they put it out under Creative Commons? Do whatever the heck you want with that license.

Now I got you. Well, yeah. So that, yeah, I think if it was licensed under do whatever you want as long as the license actually said that, then yeah, you're probably in a much better position. Yeah. Now, and I throw you another curve bowl here. Maybe this one, I got the sandpaper in my waist and I put a little scratch on it. I forgot what they called that in the big leagues,

but remember they were doing that for a while? They were scraping the famous case. One of the pitchers would scratch the ball with a little sandpaper. That's not IP, I don't know. Not IP, not your will house. Okay, you guys can look it up. It was a famous case. Just a sinker.

This would be a sinker of a pitch. All right. People are hiring humans to just, train LLMs in verticals. Okay. Now, they go and they, this is a human and they're taking notes and they've read the wire cutter and they've read consumer reports and 20 other sites and, you know, they distill it all down into their own take on it and they feed that into an LLM directly.

LLM in the clear when it does that output or not. Yeah. So I think first looking at what the human does is that a derivative work of all the content that they've read. Maybe or maybe not. The idea of it being a derivative work is it substantially similar to the original work, and then obviously by reading it, they had access to it. So you have to show access and substantial similarity to the original work. If what this person was doing is writing their own opinion on the wirecutter review and wired reviews and other reviews of content, it's more likely that that is an original work. And they are simply expressing their own opinions. But then you are training an LLM based on someone else's opinion and not the original primary source. So this is all going to go back to. And what are you using that LLM to do? Yeah, give you an answer, I think in some cases. And in other cases, make another transformative work.

I think I have an analogy for us. I remember when Metacritic and Rotten Tomatoes came out, they would create their own score. So they would read everybody else's review. And they'd say, you know, I think Roger Ebert is, you know, he doesn't like to give five stars, he doesn't like to give a number, but based on my interpretation of his review, that's a 70. That's a 90. He really loves Indiana Jones Temple of Doom, but he's not a fan of the third one. So I'm going to put a score on it based on my analysis as a, you know, cinephile. They were allowed to do it.

People were pissed off. People were upset. Sorry. We allow for people to make their, own derivative works or make their own creations based on their interpretation of what you said, right? This would be a canonical case, I guess, of fair use. Yeah, so I would say that probably fair use may not even apply because you're not copying, right? You're not doing anything by providing a numerical rating to someone's textual description of something and trying to rank it on a

scale. So yeah, I think that that is just my own evaluation or whoever that is is evaluation of, right, someone's work. Well, you know, it's really interesting about this is I think the audience from our dialogue is getting an idea of just how nuanced this is and what the lens that you can put your own behavior for your startup through. And maybe hopefully if, you know, you saw an opportunity, you could think about, well, what did Metacritic do? What did Rotten Tomatoes do? You know, what are strategies for maybe contacting the rights holder and also putting yourself in the same position. I know many technologists who really feel strongly that they should be able to do whatever they want if it's on the open web. And then when somebody in China steals their IP,

they get all bent out of shape. I encourage technologists to maybe think about that. How would you feel if your LLM was stolen, if your weights were stolen, if your code was stolen that you actually build your product from, and then somebody in another country or in ours, you know, just photocopied, website, install your design and your logo and confuse people. It wouldn't be fun, would it? Yeah. And that could be happening even unintentionally. So, depending on what AI platform you're using and what version, the terms could allow for very different things. And so sometimes a platform could say, we can take your prompts and do with it what we want. And at that point, they could train a model on your prompts. And all of a sudden, the things that you were trying to keep secret, the things you

were trying to keep proprietary could be made available to others. And so now you've sort of unknowing undermine yourself. And then the model that you're using, does it say you own the output, that you get a license to it? And so oftentimes you recommend to clients that if there are enterprise versions of whatever you're looking to use, use the enterprise version, because that is directed towards the business enterprise and will be the most protective of you and is what your investors are expecting. Awesome. Thank you so much, Chris. If you want to see all the startup basic series we've done, legal, accounting, marketing, AI, man, we've covered it all. This we can startups.com

slash basics. And if you need a fantastic attorney, use the one I use, Wilson Sincini, WSGR. Thank you, Chris. Thank you very much.