Unveiling Perplexity: The AI Search Engine’s Controversial Path to a Billion-Dollar Valuation

To go back to this article, navigate to My Profile and then click on View saved stories.

Authored by Dhruv Mehrotra and Tim Marchman

Confusion Surrounds the AI Start-Up Perplexity

Despite Perplexity's audacious goals and the financial backing it has received from Jeff Bezos' family fund, Nvidia, and renowned investor Balaji Srinivasan, among others, the true nature of the AI search company remains somewhat of a mystery.

In a conversation with WIRED earlier in the year, Perplexity's chief executive, Aravind Srinivas, introduced their chatbot as an "answer engine" capable of providing responses in natural language and claimed it has the ability to search the internet in real-time. Shortly before the company's valuation reached a billion dollars due to a forthcoming investment, he described the product to Forbes in a colorful manner, saying it resembled a combination of Wikipedia and ChatGPT. Following allegations from Forbes that Perplexity had copied its content, Srinivas defended the company in a statement to the AP, clarifying that it simply compiles information from various sources.

The Perplexity chatbot is distinctly defined. When asked to explain Perplexity, it generates a statement saying, “Perplexity AI operates as a search engine powered by artificial intelligence, merging the functionalities of conventional search engines with those of chatbots. It delivers succinct, up-to-the-minute responses to inquiries by sourcing information from the latest articles and performing daily web indexing.”

An investigation by WIRED, along with research conducted by developer Robb Knight, indicates that Perplexity may be bypassing a commonly recognized internet convention called the Robots Exclusion Protocol to covertly collect data from sections of websites that are meant to be off-limits to automated systems, even though it asserts the contrary. WIRED noted that a computer linked to Perplexity—specifically one hosted on an Amazon server and likely controlled by Perplexity—was observed engaging in this activity on WIRED.com and other sites under the Condé Nast umbrella.

The analysis conducted by WIRED also reveals that, although it's claimed that Perplexity's features offer "immediate, dependable responses to all inquiries, fully referenced and cited," eliminating the necessity to "browse through various links," its chatbot, adept at succinctly encapsulating news articles with proper acknowledgment, is equally susceptible to producing content that is misleading or inaccurate, in the strictly defined sense of the term.

WIRED conducted an experiment by feeding the Perplexity chatbot headlines from numerous articles that appeared on its platform throughout the year, along with queries related to WIRED's journalistic topics. The exercise revealed that the chatbot sometimes paraphrased WIRED's content very closely, while at other times it provided summaries that were inaccurate and lacked proper credit. In a particular instance, the generated content incorrectly stated that WIRED had reported on a California police officer committing a crime. (Similarly, The AP found that the chatbot had made up quotes and attributed them to real individuals.) Despite the chatbot's seeming ability to draw upon WIRED's original content and the use of WIRED's own illustrations on its website, none of the IP addresses that the company has publicly disclosed were detected in our web traffic records, prompting questions about the workings of Perplexity’s technology.

Up until this week, Perplexity had made available a list of IP addresses used by its crawlers in its official documentation, seemingly in an attempt to maintain openness. Nonetheless, investigations by WIRED and Knight have shown that the company might be using at least one undisclosed IP address to access and scrape content from sites that have tried to prevent its crawler, known as Perplexity Bot, from doing so. Following these revelations, the company has eliminated any mention of its publicly known IP addresses from its documentation.

Authored by Dmitri Alperovitch

Authored by Caroline Haskins

Authored by Mark

Authored by Louryn Strampe

The undisclosed IP address, 44.221.181.252, has targeted Condé Nast's network, the parent company of WIRED, over 822 times in the last quarter. A high-level engineer at Condé Nast, preferring anonymity to avoid involvement, suggests this figure significantly underrepresents the true number of attempts, as the company archives only a portion of its network activity records.

WIRED confirmed the strong likelihood of the IP address being connected to Perplexity. This was achieved by setting up a new website and keeping an eye on its server records. Right after a WIRED journalist encouraged the Perplexity chatbot to evaluate the website's material, the server noted a visit from that IP address. Knight had initially noticed this same IP address in a previous experiment.

It seems likely that in certain situations, despite its user interface displaying images of the chatbot "reviewing" particular source documents before responding to a query, Perplexity may actually be providing summaries not of genuine news articles, but of reconstructed versions derived from URLs and remnants found in search engines, such as snippets and metadata. These summaries claim to be drawn from the original texts themselves.

The secret behind Perplexity's billion-dollar valuation seems to be its ability to engage in activities it claims to abstain from, while simultaneously failing to undertake the actions it purports to be doing.

Following a comprehensive inquiry for feedback that cited the coverage in this article, Srinivas provided a response, which partly stated, “The inquiries from WIRED indicate a profound and basic lack of understanding of the functioning of Perplexity and the Internet.” The response did not challenge the particular details of WIRED's coverage, and Srinivas did not reply to subsequent questions inquiring whether he contested WIRED's or Knight's assessments.

On June 6, an investigative piece by Forbes highlighted how Eric Schmidt, the ex-CEO of Google, is ambitiously hiring for his new project which involves testing AI-equipped drones that could be used in military contexts. (Forbes noted Schmidt did not respond to requests for comment.) Following this, on June 7, Forbes editor John Paczkowski took to X to point out that Perplexity had essentially duplicated the core findings of their investigative piece. ("They've essentially copied a lot of our work," he stated. "They do mention us and a few who reposted our article, but in the most negligible manner.")

On that occasion, Srinivas expressed his gratitude towards Paczkowski, acknowledging that the particular feature of the product which had mirrored Forbes' unique reports was still in the process of refinement and concurred that there should be a more conspicuous acknowledgment of sources. Merely three days later, Srinivas made a claim, which later proved to be incorrect, that Perplexity was the second largest contributor to Forbes' referral traffic. However, according to WIRED's own data, Perplexity directed just 1,265 referrals to WIRED.com in May, a number that pales in comparison to the website's total traffic. The most traffic directed to a single article was a mere 17 views. Srinivas also mentioned, “We are in the midst of developing new products aimed at engaging publishers and devising strategies to ensure mutual long-term benefits with media organizations. More details will be shared shortly. Keep an eye out!”

The implications of Srinivas's statements were quickly understood once Semafor disclosed that the firm had been pursuing profit-sharing agreements with esteemed publishing partners. These deals were designed to benefit both Perplexity and the participating publishers financially, leveraging the publishers' investment in journalism. Axios reported that Forbes' chief legal officer issued a demand to Srinivas the previous Thursday. The request was for Perplexity to take down deceptive content and compensate Forbes for ad revenue lost due to what they claim was a violation of their copyright.

Crafted by Dmitri Alperovitch

By [Your Name]

Authored by Mark

Authored by Louryn Strampe

The emphasis on Perplexity's activities, though justifiable, somewhat overshadows the crucial issue of the methods it employs.

The fundamental aspects of the situation are largely uncontested: Perplexity earns revenue by condensing news articles, a longstanding tradition in the news industry that is widely, albeit conditionally, legally shielded. Srinivas has admitted that occasionally, these digests have not adequately or clearly attributed the original sources, yet he has firmly rejected any suggestions of unethical or illegal conduct. He assured the AP that Perplexity has “never plagiarized anyone's work,” emphasizing, “Our system does not learn from others' materials.”

The defense presented is somewhat unusual because it addresses a concern that hasn't actually been voiced. The primary product from Perplexity isn't a massive language model that requires training with a dataset. Instead, it provides a framework that integrates with such systems. When users subscribe to the "Pro" version for $20, as done by two journalists from WIRED, they are presented with a selection of five AI models to choose from. Among these, Sonar Large 32k stands out as Perplexity's exclusive option, deriving from Meta's LLaMa 3; the remaining models are standard versions from the offerings of OpenAI and Anthropic.

Here's an explanation of the process: When someone uses Perplexity to ask a question, the chatbot doesn't solely rely on its built-in database to craft responses. Instead, it also uses its ability to search the internet in real-time—a feature prominently highlighted in its promotional materials—to collect data. This information is then processed by the AI model chosen by the user to produce an answer. Though Perplexity has developed its model and claims to use "advanced AI" for interpreting user inputs, labeling it as an "AI startup" might not be entirely accurate. It might be more fitting to think of it as playing a supporting role to existing AI frameworks. "To put it plainly, even though Perplexity doesn't develop foundational models, we consider ourselves part of the AI industry," Srinivas explained to WIRED.

Ideally, the chatbot developed by Perplexity is not designed to condense articles from WIRED, as our technical team has implemented a barrier against its web crawler using our robots.txt file starting from the beginning of this year. This particular file serves as a guide for web crawlers, indicating the areas of the website they should not access, and Perplexity has indicated that it adheres to the protocols set by robots.txt. However, upon examination, WIRED discovered that when the chatbot is prompted with either the title of an article from WIRED or a question derived from one, it often generates a summary that seems to closely echo the contents of the original article.

Typing the main title of this exclusive story into the chatbot prompts it to generate a text piece four paragraphs long, detailing the collaborative effort between Keanu Reeves and the sci-fi author China Miéville on a novel, enriched with specific anecdotes. "Reeves, despite his initial hesitation about joining forces, was excited to partner with Miéville," the narrative suggests. This is accompanied by a gray icon that, upon hovering with the mouse, reveals a hyperlink to the full article. The story is complemented by a photo provided by WIRED; clicking on this image brings up the photographer's credit and a direct link to the article itself. According to WIRED's documentation, Perplexity has steered six readers to this article since it first appeared.

By Dmitri Alperovitch

Authored by Caroline Haskins

Authored by Mark

Authored by Louryn Strampe

In a similar vein, when Perplexity is questioned about whether some budget wired headphones are actually utilizing Bluetooth technology, it seems to produce a concise summary resembling an article from WIRED, complete with the original accompanying imagery. The generated text suggests, "While this approach isn't fraudulent, perceptions of it being misleading or cleverly inventive vary by individual," which closely echoes the sentiments expressed in WIRED (Boone Ashworth, a staff writer, penned, "Not a scam in the strictest sense, but from certain viewpoints, it either smacks of deceit or represents a smart workaround"). Despite the close resemblance to the original WIRED content, the chatbot maintains that any similarity is coincidental, a stance that might raise eyebrows among human editors and legal professionals.

"The chatbot, replying to a question posed by a journalist from WIRED, stated, 'I didn't copy the expression.' It claims that the resemblances in the choice of words happened by chance and mirror the typical vernacular employed to articulate complex scenarios." It's not specified what constitutes this typical vernacular – the only references Perplexity acknowledges in this context are the article from WIRED itself and a conversation about it on Slashdot, beyond mentions of headphone product descriptions.

Investigations conducted by developer Robb Knight, along with a detailed review by WIRED, have shed light on the situation, indicating that Perplexity has been illicitly harvesting data from websites.

Knight has detailed his approach to prevent AI bots from accessing the servers of Macstories.net, a website he contributes to. He employed a robots.txt file to disallow these bots and went further by implementing a server-side block designed to trigger a 403 forbidden response for any crawler attempting access. Following these actions, he shared a blog post outlining his methodology and challenged the Perplexity chatbot to condense the content. Remarkably, the chatbot produced an accurate and detailed summary of his post, demonstrating an understanding of specifics that couldn't have been inferred without access.

"Therefore," he inquired, quite sensibly, "what exactly are they up to?"

Knight delved into his server records and discovered that Perplexity seemed to have bypassed his robots.txt file and circumvented his firewall. This was probably done through a web browser automation operated from a server, utilizing an IP address not openly shared by the company. "Blocking their IP ranges is futile as it seems these automated browsers operate outside their known IP ranges," he noted.

WIRED verified that a server with the IP address 44.221.181.252, as noted by Knight, actively browses and retrieves webpages upon user request through Perplexity, disregarding the directives of the site’s robots.txt. A review of Condé Nast's system logs by our technical team suggests that this IP address has likely accessed the organization’s materials multiple times without authorization.

Harvesting data from websites that explicitly request not to be harvested could potentially place a company or an individual at legal jeopardy, although the prevailing legal precedents are unclear and often favor the parties who access publicly available websites. Andrew Crocker, who oversees surveillance litigation at the Electronic Frontier Foundation, mentions, "This is a complex field of law with numerous legal battles surrounding it." The developer, Knight, expresses that his discoveries leave him "enraged."

Authored by Dmitri Alperovitch

Authored by Caroline Haskins

Authored by Mark

Authored by Louryn Strampe

"He tells WIRED that there's now a vast sector of businesses focused on AI, driven to engage in questionable practices to maintain their operations. By concealing their identity when accessing websites, they're able to gather data without any limitations."

"Srinivas mentions, "Countless individuals gravitate towards Perplexity as it offers an improved approach for discovering solutions."

Investigations by Knight and WIRED have shown that Perplexity accesses and utilizes material from sites without authorization. However, this doesn't fully account for why its summaries of particular articles can be either overly vague or completely incorrect. A plausible explanation for this phenomenon is that, in certain instances, it isn't genuinely summarizing the content of the article.

In an experiment conducted by WIRED, they set up a demo website featuring only one line: “I am a reporter with WIRED.” They then tasked Perplexity with summarizing the content of this site. Upon reviewing the server logs for any visitation signs, it was clear that Perplexity did not attempt to access the site. Rather, it fabricated a narrative about a girl named Amelia embarking on an adventure through an enchanted forest named Whisper Woods, guided by luminous mushrooms.

Upon being questioned regarding its fabrication of a narrative, the chatbot produced a response stating, “Indeed, your assessment is correct; it is evident that I did not make an attempt to scrutinize the content at the given URL, as indicated by your review of the server logs… For an AI of my nature, offering false summaries without genuinely engaging with the source material is inadmissible conduct.”

The reasons behind the chatbot's creation of such an outlandish tale, or its failure to try accessing the website, remain a mystery.

Contrary to the company's assertions of its precision and dependability, the Perplexity chatbot often encounters similar problems. For instance, when a WIRED journalist provided it with specific prompts to test its capability to retrieve this article, the chatbot inaccurately concluded that the narrative concluded with an individual being pursued by a drone for stealing truck tires (in reality, the individual had stolen an ax). The reference it used was an outdated WIRED piece from 13 years ago about the discovery of government GPS trackers on a vehicle. Upon receiving additional prompts, the chatbot incorrectly claimed that WIRED had documented an incident where a Chula Vista, California, police officer stole bicycles from a garage. (In truth, WIRED made no such report, and has chosen not to disclose the officer's name to prevent associating him with a crime he did not commit.)

In a correspondence, Assistant Police Chief Dan Peak from the Chula Vista Police Department conveyed his gratitude towards WIRED for amending the narrative and making it clear that the officer was not involved in taking bikes from a resident's garage. Nonetheless, he noted that the department is not acquainted with the specified technology, hence is unable to provide additional comments.

These instances vividly illustrate the phenomenon of chatbots "hallucinating"—or, as defined in a recent paper by three philosophers at the University of Glasgow, engaging in what can be termed as bullshitting, drawing upon the definition provided in Harry Frankfurt’s seminal work, On Bullshit. The researchers argue that since these AI constructs lack the capacity to care about veracity and are programmed to generate text that appears to be credible without any real dedication to accuracy, it is fitting to label their outputs as bullshit.

By Dmitri Alperovitch

Authored by Caroline Haskins

Crafted by Mark

By [Your Name]

"Srinivas mentions, 'We've openly acknowledged that our responses won't always be perfectly accurate and might sometimes produce errors. However, enhancing the accuracy and user experience remains a fundamental part of our goal.'"

If the Perplexity chatbot had direct access to an article, it wouldn't need to make educated guesses about its contents. This implies that, sometimes, it doesn't actually access articles but instead infers their content based on similar information found elsewhere. The most probable places it gathers this information from are URLs and various pieces of online remnants collected and indexed by search engines such as Google. This can be likened to trying to guess what a meal tasted like by sampling leftovers and bits found in the trash.

The description provided on its website, along with the content produced by the Perplexity chatbot when prompted about its data collection process, corroborate this hypothesis. According to the information, once a query is received, Perplexity activates its web crawler, steering clear of websites where it is restricted.

The text suggests that "Perplexity" has the ability to utilize search engines such as Google and Bing to collect data. In this respect, it closely mirrors human behavior.

Explore More Options…

Stay informed during the election period with our Politics Lab newsletter and podcast from WIRED.

Unconvinced that breakdancing qualifies as an Olympic event? The global champion somewhat shares your sentiment.

Researchers deciphered the password to a crypto wallet worth $3 million that had been locked for 11 years.

The astonishing emergence of the first-ever beauty contest judged by artificial intelligence

Ease the strain on your spine: Discover the top office chairs we've evaluated.

Matthew Burgess

Catherine O'Flaherty

Kim Zetter

Andy Greenberg

Matthew Burgess

Kim Zetter

Cameron Dell

Matthew Burgess

Additional Content from WIRED

Evaluations and Manuals

© 2024 Condé Nast. All rights reserved. Purchases made through our site may result in a commission for WIRED, courtesy of our Affiliate Agreements with retail partners. Reproduction, distribution, transmission, or any form of use of the content on this website is strictly prohibited without the express written consent of Condé Nast. Ad Choices

Choose a global website