Perplexity CEO Champions AI Agents, Warns OpenAI Chrome Control Jeopardizes Open Web

Aravind Srinivas argues that agents must gain direct entry to smartphone apps to fulfill their potential, and he warns that giving OpenAI control of Chrome would weaken a web held together by open source software.

Perplexity, a search venture powered by generative language models, rose to prominence during 2024 as a contender to Google’s search engine. Its conversational interface has drawn praise from people seeking context and explanation instead of simple link lists, but the company has also been sued for alleged copyright breaches. Critics, including Forbes, say that Perplexity has scraped articles, lightly paraphrased other sites, and sometimes delivered inaccuracies.

In spite of that controversy, Perplexity says it handles about 650 million queries each month and is pursuing funding at an $18 billion valuation. The team is expanding into phone assistants and is building its own browser, aiming to embed AI more deeply in how people find and interact with information.

In April, Motorola announced that its new Razr Ultra models will ship with Perplexity preinstalled. Last month, Perplexity struck a deal with PayPal to let users complete purchases via its assistant. Bloomberg reports that Samsung may follow suit by bundling the technology on its devices, though the startup declined to comment on that.

CEO Aravind Srinivas spoke with WIRED senior writer Will Knight by phone and email. The exchange below has been edited for clarity and brevity.

WIRED: The PayPal deal seems important to the vision that everyone has for agents. Is ecommerce the killer “agentic” app?
Aravind Srinivas: "Agents are the killer app for everything. Agents allow users to have the experience that’s best for them. Some people like shopping and researching, and some people want it done for them. There’s a spectrum in between, and our focus is what is the best experience for the user."

He notes that agents can adjust their behavior to match a user’s preferences, handling everything from in-depth research to fully hands-off purchasing.

WIRED: Speaking of the experience, agents still make mistakes. What happens when they buy something by accident?
Aravind Srinivas: "Merchants and buyers have adapted to every new technology since ancient times. We just show both of them what is possible, and they choose. Every successful technology has had to take security and error-resolution very seriously, and that won’t change."

He argues that dispute-resolution systems and refund processes have evolved alongside each new innovation, so accidental transactions will be handled without derailing user trust.

WIRED: Integrating AI with personal devices is another big theme. Why does the Motorola deal matter to you?
Aravind Srinivas: "It’s a big deal because Motorola is one of the largest phone brands in the world. This partnership gives us the ability to make trustworthy AI more accessible than ever. Now, by introducing Perplexity to millions of people around the world, in a native and seamless way, more people will get to see how much more really is possible with search."

WIRED: Would you consider developing your own devices eventually?
Aravind Srinivas: "We are focused on building the best AI assistant and answer engine."

He emphasizes that the priority remains delivering accurate responses and reliable automation rather than designing hardware.

WIRED: Motorola will offer other AI assistants, so how will Perplexity distinguish itself?
Aravind Srinivas: "As AI assistants become more common, accuracy and trust will become even more important. An assistant isn’t very useful if it’s unreliable. Worse, if an assistant is misleading or sycophantic, then it isn’t an assistant—it’s a manipulator. That’s not just useless, it’s dangerous. Inaccurate AI has a negative compounding effect, and we have always been the leader in developing AI and AI assistants focused on accuracy and verifiability. That will have a positive compounding effect."

He warns that an assistant must not simply echo the user’s biases. Instead, it should provide verified information, since mistakes can multiply through automated workflows.

WIRED: Wait though … Perplexity—like other AI search engines—has been criticized for hallucinating and getting things wrong.
Aravind Srinivas: "We welcome this criticism, because it’s the best way for us to continually improve. In reality, errors account for a small fraction of results, and our answers are far more accurate than 10 blue links polluted by decades of SEO-optimized content. But the fact is, accuracy and trust will only become more important as AI integrates into more of our lives, so this is something we’re relentlessly focused on. We can’t get there without this feedback."

He adds that ongoing user feedback and corrections are critical to raising the bar on reliability, and he claims Perplexity already outperforms traditional search engine results in relevance.

WIRED: Perplexity also cribs from copyrighted news articles with its “discover” section. Do you understand why some publishers are upset?
Aravind Srinivas: "We’ve answered that before. See our blog post on how we respect robots.txt."

He points to compliance with robots.txt files as part of the company’s strategy to access publicly available content while honoring publisher rules.

WIRED: The Perplexity assistant for Android and iOS seems “agentic” because it can take actions. How big of a shift is this?
Aravind Srinivas: "AI is pretty good at answering questions now. What really needs to be done is get AI to take actions. People use the word “agents”; you can go with whatever you want—“agent” or “assistant”—but in the end, it needs to string together tools and execute actions. That’s why we’re [also building a] browser, and an assistant on iOS, Android."

WIRED: Do Apple and Google have too much control over their mobile platforms compared to outsiders looking to build agents?
Aravind Srinivas: "With iOS it’s particularly challenging, because you have to string together a bunch of event APIs. On iOS, Mail, Calendar, Reminders, Podcasts, all that stuff is natively available through the Apple SDK [software development kit used to build applications], so you can actually at least draft emails, schedule meetings, move meetings, set reminders, all this stuff, open podcasts pretty easily. You can do searches for podcasts … “get me the one where Mark Andreessen discusses de-banking with Joe Rogan.” It can get you that pretty quickly."

Aravind Srinivas: "It’s mostly difficult because you cannot access other apps. iOS is not very different from Android, because AI cannot access most apps on Android either. But third-party apps can build their SDKs to be accessible on the Android SDK. For example, our Android system can display a song on Spotify. On iOS, you can only link to a specific Spotify song, and you have to manually start playing the audio."

He stresses that app developers hold the key by exposing APIs that allow AI to interface directly without forcing a user to open the app.

WIRED: Oh, so it’s app-makers that are holding AI agents back?
Aravind Srinivas: "That’s the challenge. If people are offering us APIs—say, Open Table, Uber, DoorDash, or Instacart—where we can access information within the app without even having to open the app. On the back end, that’s pretty powerful. For example, if we can access information on Uber and find that Uber comfort doesn’t cost more than 5 or 10 percent of Uber X, then we can just book Uber comfort for you—if that’s a preference that you set on Perplexity."

Aravind Srinivas: "Or similarly [we can] find the best Thai place near you and get me a dash [delivery] a lot faster than going to DoorDash app, searching for Thai food and scrolling through all these options, reading all those reviews and then putting your address, doing the checkout, all that stuff. We could honestly do all that in our system and just make the experience a lot more seamless and simple. I think that’s where things are headed, but people need to open up their apps to us, and we’ll have to see who’s willing to do what."

He believes an integrated assistant could collapse multiple steps in a food or ride service into a single command, but it will require cooperation from service providers.

WIRED: Isn’t the biggest problem that AI agents just aren’t very smart and useful yet?
Aravind Srinivas: "My analogy [for AI agents] is that we are at a point where Perplexity was in 2022 [just before it took off]. It’s not like we got all the answers right, people made fun of hallucinations and some people call it “Google in macro seconds.” It was not quite there."

Aravind Srinivas: "It only took off many months later, when models got better, and I expect the same trend with the agents and assistants. There’s going to be some set of things that really work, daily use cases, and there’ll be a long tail of things that don’t work, that we’re going to keep fixing over time."

Aravind Srinivas: "But that’s exactly why we are building a [web] browser, because the browser front end will let you do the work on your own too, if you’re not happy with what the AI did. So, then we can learn from that and fix that over time. Waymo and Tesla self driving did not work for a long time. Now, people take them for granted. I think it is a similar trajectory for us."

He likens the evolution of agents to the rollout of self-driving cars—early setbacks give way to widespread acceptance once maturity is reached.

WIRED: Is this why you floated the idea of Perxplexity taking control of Chrome—if Google were forced to divest it?
Aravind Srinivas: "We’re not saying we’re interested in buying Chrome. We’re saying that if there’s no other path, if Google is put in a situation where they have to divest Chrome, then we’d be open to running it. But Google should not have to divest Chrome, because Chrome and Chromium are tied to each other."

Aravind Srinivas: "OpenAI has also shown an interest in taking control of Chrome."

Aravind Srinivas: "Giving ownership of Chrome or Chromium to a company like OpenAI would be a disaster, because open source and OpenAI are an oxymoron at this point."

Aravind Srinivas: "There are only two companies that can really potentially run Chrome: Microsoft and Meta. Honestly, Microsoft would spoil it, just like it spoiled Edge. And transferring Chrome to Meta is transferring Chrome from one monopoly to another."

He says regulators should consider any browser change carefully, since handing the project to one company risks shifting a monopoly from one gatekeeper to another.

WIRED: What do you expect agents to be useful for first?
Aravind Srinivas: "I think [they will make] a lot of your personal searches a lot better."

Aravind Srinivas: "Like asking, “What was the article I read last week about this one particular company?” or “Can you summarize my X feed for me so that I know what’s trending?” because you don’t want to go to X and get distracted by it. Or “Can you schedule this meeting for me with this person and if there’s a conflict send them an email asking for a different time?” All these boring things, I feel we will be able to automate [them] pretty quickly [in future]."