As one of the most visible AI businesses in the world, Perplexity faces legal questions on several fronts. In this episode of AI Decoded, Proskauer partner Oliver Howley and Nathan Barksdale, General Counsel and VP of Operations at Perplexity, discuss AI litigation, global regulation, and common misconceptions in AI contracting.
OLIVER HOWLEY: I want to discuss something that is not discussed publicly often. As one of the most visible AI businesses in the world, you inevitably attract the attention of the plaintiffs’ bar. And my team has worked with you on a number of matters. But I just want to discuss openly how you think about litigation strategically, and how do you actually decide which fights are worth having?
NATHAN BARKSDALE: It’s not us necessarily deciding the fight — we are deciding to defend our rights and defend users’ access to information. It’s human nature, to any new developments, new technology, any disruptive industry, to have a lot of resistance against it. It’s not new to us, we’re not surprised. In my personal point of view, I think any industry that has resisted change based on new technologies has a negative outlook. We’re seeing it within the publisher community that those publishers that are most open, I would bet that those are going to be the publishers that will thrive the most into the future, because they are adapting the technology, they are adjusting their business models, and they are not looking at it as “This is impeding on our bottom line, and how do we stop it,” right?
OLIVER HOWLEY: So you’re subject to a number of suits at the moment — we’re working with you on a number of them — claimants ranging from Dow Jones through to Encyclopedia Britannica through to Chicago Tribune. What is the most frustrating thing for you in those claims? What are the most frustrating arguments that you’re seeing?
NATHAN BARKSDALE: For the most part, we’re finding that there’s a lot of pushing our answer engine to deliver results in a way that it wasn’t intended to. Most of that is regurgitating copyrightable articles of the entire breadth of the article, or information that the engine doesn’t necessarily want to provide, but it’s being tried again and again, in different ways, to do so. Eventually it will break, right — they will find a way around it, and we can patch that up. But there are going to be other actors that do the same thing, and we find that they’re bringing suits based off of what is the last result that was developed, but not showing the entire chain of events where our system was designed not to regurgitate and provide that copyrighted information.
OLIVER HOWLEY: So they’re pushing and pushing and pushing for the purpose of manufacturing the claim?
NATHAN BARKSDALE: Right.
OLIVER HOWLEY: But what can you do as a business to try and counteract that, to stop that?
NATHAN BARKSDALE: We can continue to set up guardrails and learn from what’s happening in those instances to make sure that it doesn’t happen again. Companies said they don’t want their article cited, and that was a big issue, even though we were lawfully able to retrieve information and summarize it in a way that was legal — we made sure that we were citing more sources, so that you wouldn’t see one publisher or one source cited in an answer; instead you receive a summary that’s a compilation of many different sources and factors. And then we can also make sure that we are abiding by robots.txt entirely, which we have always done, and we will ensure that we continue that practice going forward.
OLIVER HOWLEY: In a Bloomberg interview recently, you described many of the lawsuits brought against Perplexity as reflecting a, I think in quotes, “technological misunderstanding.”
NATHAN BARKSDALE: Right.
OLIVER HOWLEY: What did you mean by that?
NATHAN BARKSDALE: Yeah, I think when we look at all of the AI suits in the industry, the public — or many of the plaintiffs — they’re kind of lumping it together. They’re lumping us together with OpenAI and Anthropic, a lot of the foundation labs, on what they’re doing, and saying it’s an AI-related suit. It’s copyright — “Let’s sue them because they’re using our content.”
OLIVER HOWLEY: Training. Ripping us off. All of this stuff.
NATHAN BARKSDALE: Exactly. Yeah. But we’re not doing that. We don’t train foundation models. We are simply retrieving information that’s publicly available on the internet. We are summarizing it and taking snippets, and we have been refined to make sure that we’re taking no more than is necessary to provide factual information back based on a variety of articles.
OLIVER HOWLEY: That the user has asked for as well.
NATHAN BARKSDALE: Yes. So this morning, for example, I asked, “What is the top news story of the day?” And I received a line back saying the top news story is the escalating conflict between the U.S. and Iran, and it gave me three bullet points. And it cited every single line of where that information came from. Usually it’s based off of — this citation is being the most credible source, and you want to be a credible source, you want to be cited, because a lot of that factual information could be seen across many different articles. But now, us citing each of those lines has now worked to our detriment, because website owners across the internet are saying “I didn’t give you permission to use my content,” and they’re looking at that as “you’re using it,” and it’s not necessarily dispositive of where that information came from. It’s not necessarily that we scraped their website if they invoked robots.txt protocols for us not to do that. But now there’s an understanding that, “well, that’s not okay — you can’t summarize, you can’t cite facts, and you can’t display information in that way because that’s our copyrightable content.”
OLIVER HOWLEY: It’s a change, isn’t it? Because years gone by and people were always open to some summarization with a citation. And—
NATHAN BARKSDALE: Right.
OLIVER HOWLEY: And “This came from this person please go and have a look at their full article,” and that’s exactly what you do. But now there seems to be this, you know, I would say emerging, but pretty forceful pushback which seems to reflect a shift in thinking that it doesn’t really, to my mind, make too much logical sense.
Naturally, that brings us on to regulation, and there is a split of approaches when it comes to regulation around the world. The EU is almost fixated on regulating, perhaps for the sake of regulating, albeit recognizing now that it might need to walk itself back for interests of competitiveness of EU versus the U.S. and China. You’ve got the U.S., which on a whole is avoiding regulation apart from in certain states like California. You’ve got the UK in the middle. How do you see those differences impacting you as a business, your growth, your focus, and perhaps the wider AI industry and its prosperity around the world?
NATHAN BARKSDALE: For Perplexity itself, the regulation that has been proposed hasn’t impacted us too much, I think, because we are not training foundation models — we’re integrating these models, so our regulatory compliance requirements are a little bit less than say some of the other big —
OLIVER HOWLEY: —relative to the other big ones.
NATHAN BARKSDALE: Exactly. I see the landscape between the U.S. as you describe, the UK, EU, and then APAC. And what I’m seeing at a higher level is how each country and business segment’s use of AI is going to be impacted by regulatory measures in those three particular regions. In the U.S., being very low regulation at the moment — there’s a lot of proposals out there, there’s a lot of room for dialogue. We welcome communication with AGs and congressional officials, and I feel like there is a time in the U.S. to develop regulation that makes sense here. In the EU, we saw a pullback with the EU AI Act, and I think that was warranted — I commend them for doing that instead of charging ahead on what they originally thought was the right decision. And in APAC, it’s interesting, because you have China dominating with their models, and you have a lot of other countries in APAC that seemingly don’t want to be left out, but they are more prone to regulation, and so how they balance those two things — welcoming offshore foreign models and products like Perplexity into their regions, but also being very cautious — they know they can’t get left behind, and that’s where I see a lot of things taking place, because they want to regulate, but they know that it will be detrimental to their forward progress if they do so. That’s something we’re watching very closely.
OLIVER HOWLEY: You and I have negotiated from both sides of the table for the business, so, as a buyer and as a vendor, you and I have these debates all the time on what matters, what doesn’t matter. What’s the biggest misconception that customers seem to have when they’re negotiating AI contracts, deals with you? Is there something – and I have a view on this – that procurement teams seem to obsess over or fixate on unnecessarily?
NATHAN BARKSDALE: Ultimately, I think it’s that security and compliance does not apply to AI companies, and that AI companies can do whatever with the data and information they have access to — that anyone in the organization can access and control it, that your data is out in the cloud being accessed by multiple different parties in ways that —
OLIVER HOWLEY: All of the other customers of the same. [Cross talk] Yeah, yeah.
NATHAN BARKSDALE: Exactly, and that’s simply not the case. The same rules that apply to Google, to Amazon, to Snowflake apply to all of the AI companies. We do not train, but we’re constantly having to inform all of our customers that we do not use your data to train. But it doesn’t mean that downstream — the OpenAIs, the Anthropics, the models that we use — that they have that same understanding. So we make that clear to all of our customers —
OLIVER HOWLEY: Sure.
NATHAN BARKSDALE: — that none of our sub-processors will train on their data and information, but there is still this kind of doubt that still is taking place. And I think we as an industry need to overcome a lot of those talking points, to get people to understand that we are protecting your data. We have internal controls to protect your data as well, and that downstream sub-processors and other information is not used in any other way than to provide the services to you.
OLIVER HOWLEY: I think it’s a bit unfortunate, isn’t it? Because there seems to be a fixation on the AI industry and the mischaracterization that you will all misuse customer data. And perhaps that’s driven by recent suits and recent, you know, disclosures around certain AI businesses, misusing printed books and publications, and customer data for their own means. But it seems to be tarnishing the entire industry rather than just individual vendors.
NATHAN BARKSDALE: And I think people are right to question because the technology is so powerful. And there’s so much access to information and with stronger models and stronger capabilities of companies, you’re going to going to get companies and information or people providing more information to those systems—
OLIVER HOWLEY: Sure, sure.
NATHAN BARKSDALE: —And they have a right to ask questions, but it doesn’t help companies like Perplexity to misuse data because we ultimately want to provide a service, we want to provide a service to our customers, and the only way they feel comfortable using us is if they feel like their data is secure.
From the beginning when we have those customer conversations, there is almost a distrust. Like we do not trust any companies, we have to use you, and we want to one, be able to combat that, but how do we kind of reset the conversation and narrative so where that isn’t the first conversation step taking place. And I think, you know, making sure that our policies are clear, that information is out there. And we’re always thinking more innovative about how to do that. And I think we have a little bit of a ways to go to get people to be on that same page.
OLIVER HOWLEY: It’s interesting, I often get asked by clients when I’m acting on the customer side, because obviously act for developers a lot of the time. But customers sometimes saying how can you ensure that there is no misuse of my data? Are you sure? And the narrative that, or at least the answer that I give is number one, I act for developers, and I know that they don’t misuse them. And I can say that with a straight face.
But also, if there is misuse of data at an enterprise level for an entire customer base, it would be the end of the business if it ever came out, right?
NATHAN BARKSDALE: So what we hope to come with it is stronger regulation, stronger systems, and additional compliance mechanisms that would help that — how can people feel comfortable engaging and using our products immediately, without the burden of overcoming all of the questions that typical industries have already dealt with for years.
OLIVER HOWLEY: Yeah, and I think that will be important because, understandably, you can’t allow people to audit your business, right? You can’t allow people to come in, look under the hood, come to San Fran and see what’s going on in the data centers, how you’re using the data. So those additional measures will be super important going forward.
NATHAN BARKSDALE: Right.