Why is Perplexity controversial?

The Lawsuits: Why Publishers Are Suing Perplexity

Why is a company valued at up to $20 billion spending so much time in courtrooms? For Perplexity AI, the answer lies in an escalating stack of copyright litigation from some of the world's most powerful publishers — all of whom allege the same core problem: Perplexity builds its answers on their content without permission.

The most prominent case comes from Dow Jones and the New York Post, whose suit alleges "a massive amount of illegal copying" into an internal database, at times reproducing content verbatim. News Corp CEO Robert Thomson called it "an abuse of intellectual property that harms journalists, writers, publishers and News Corp." The suit also alleges Perplexity attributed hallucinated quotes — including one about F-16 jets for Ukraine — to articles that never contained them.

The litigation has spread globally. Japan's Yomiuri Shimbun alleges "free-riding" on 120,000 articles between February and June 2025, joined by Asahi Shimbun and Nikkei. The New York Times sent a cease-and-desist in October 2024, the BBC issued a legal threat in June 2025, and Reddit filed a federal suit in October 2025, according to case documentation.

The publishers' core arguments fall into a few recurring themes:

• Competing for the same audiences using copyrighted content, without licensing agreements
• Copying "vast" quantities of publisher work into retrieval databases, sometimes verbatim
• Attributing fabricated quotes to real articles, damaging publishers' credibility

Perplexity's defense is consistent: CEO Aravind Srinivas says the company "meticulously cites sources and merely curates existing online content," comparing the practice to journalistic curation, while acknowledging "rough edges" in a Forbes case where paywalled reporting appeared in Perplexity's Pages tool with minimal attribution. The company also launched a revenue-sharing Publishers Program in July 2024, offering a double-digit percentage, flat-rate, per-article share with partners including Fortune, Time, and Der Spiegel.

It's worth stressing that these lawsuit claims remain allegations from filings, not verified facts. Licensing startup TollBit notes Perplexity is "not alone among AI companies in circumventing a common web standard" — and News Corp itself struck a multi-year OpenAI licensing deal in May 2024, suggesting publishers will partner when terms are acceptable.

For anyone building AI workflows on top of tools like Perplexity, the saga underscores why provenance and verification matter. SkillzBench is building exactly that kind of transparency for agent skills — commit-SHA pinning, security scans, and license checks — so you can see what you're actually running, not just what a vendor claims.

Stealth Crawling: What Wired and Cloudflare Found

When a website posts a "no trespassing" sign, most bots obey. Two independent investigations a year apart allege Perplexity does the opposite — disguising its crawlers to slip past the very fences publishers build.

The first investigation landed in June 2024, when Wired and web developer Robb Knight reported that Perplexity appeared to use undisclosed IP addresses and spoofed user-agent strings to bypass robots.txt files and web application firewalls — despite the company's public claims that it respects those standards, per documented reporting on the controversy. The numbers were hard to dismiss. Wired found that an IP address "almost certainly linked to Perplexity and not listed in its public IP range" visited its parent company's sites more than 800 times in three months, according to CNBC's coverage.

A year later, the allegations resurfaced from an unexpected corner. In August 2025, Cloudflare — the infrastructure giant that powers a huge share of the web — independently reached similar findings about Perplexity's crawling behavior. The rhetoric turned sharp on both sides:

• Cloudflare CEO Matthew Prince said Perplexity acts "more like North Korean hackers" than a reputable AI company.
• Perplexity fired back, calling Cloudflare's research a "charlatan publicity stunt."
• CEO Aravind Srinivas previously claimed Perplexity doesn't ignore robots.txt and blamed third-party crawlers — while declining to commit to stop scraping Wired's content through those third parties.

The company's shifting explanations have fueled much of the reputational damage. Srinivas has defended Perplexity as merely curating content "akin to journalistic practice," while acknowledging "rough edges" in a separate plagiarism dispute with Forbes, as industry reporting notes.

There is also an important mitigating fact. Per the licensing startup TollBit, Perplexity is "not alone among AI companies in circumventing a common web standard" — it has simply drawn the most scrutiny, according to CNBC. That context matters: the stealth-crawling problem is arguably an industry-wide pattern, with Perplexity as its most visible flashpoint.

For anyone building AI workflows, the episode is a reminder that provenance and verification can't be taken on faith — the same reason SkillzBench pins every agent skill to a commit SHA and security-scans it before benchmarking. Whether these crawling allegations hold up legally remains open; what's documented is that two independent parties, a year apart, reached strikingly similar conclusions.

Bias and Hallucinations: When AI Search Gets It Wrong

An AI search engine that puts words in publishers' mouths is arguably more dangerous than one that just copies them. That is exactly what Dow Jones and the New York Post allege in their lawsuit against Perplexity: the platform attributed hallucinated quotes — including fabricated remarks about F-16 jets for Ukraine — to real articles that never contained them, according to court filings summarized on Wikipedia. The publishers' work becomes a coat rack for invented content, and their brands absorb the damage.

The bias picture is just as concerning for anyone relying on AI search to inform decisions. A July 2025 arXiv preprint analyzed 24,000+ conversations, 65,000 responses, and 366,000+ citations across Perplexity, OpenAI, and Google. It found news citations concentrate heavily among a small number of outlets and display a "pronounced liberal bias" — though low-credibility sources were rarely cited. Study author Kai-Cheng Yang warns these systems are "emerging as new information gatekeepers."

What makes the bias finding especially troubling is how little users appear to correct for it. The same study found that user satisfaction was not significantly influenced by political leaning or source quality, implying readers may not notice when the gate is tilted. With at least 46% of Americans now reporting they use AI tools for information seeking, the stakes extend well beyond academic curiosity.

Hallucinations, meanwhile, may be a feature of the architecture rather than a bug. Harvard Kennedy School researchers argue AI inaccuracies are "structurally inevitable" — the product of biased training data, opaque training, and weak downstream gatekeeping. Even retrieval-augmented systems like Perplexity, which ground answers in live sources, can still fail through conflicting sources or "poisoned retrieval." And because AI outputs carry a fluent, authoritative tone, the researchers note they can be more persuasive than human-written misinformation.

For business readers, the practical risk looks like this:

• Fabricated quotes attributed to real, trusted sources — the Dow Jones allegation — can slip into market research or competitive briefings undetected.
• Citation concentration means a small set of outlets shapes what AI search surfaces, narrowing the range of perspectives feeding your analysis.
• Bias that users don't detect, per the arXiv findings, won't self-correct in the answers you read.
• Authoritative tone plus structural inevitability means fluency is not evidence of accuracy.

None of this means AI search is unusable — hallucinated academic references dropped to roughly 0.6% for ChatGPT on well-established scientific topics in one 2025 study. But it does mean verification has to be deliberate, not assumed. That principle drives how SkillzBench is being built: pinning agent skills to exact commit versions, security-scanning, and benchmarking each one in a sandbox before it earns a place in its free directory. Whether the output is a search answer or an agent's action, provenance and independent testing are the only reliable guardrails.

How the Industry Is Responding: Suits vs. Revenue Sharing

The AI industry has split into two camps: publishers taking Perplexity to court, and publishers taking Perplexity's money. Which track a media company chooses depends largely on whether it believes an answer engine can coexist with the people who create its answers.

The litigation track is growing. Dow Jones and the New York Post sued Perplexity in October 2024, alleging a "massive amount of illegal copying" and seeking a court order to destroy internal databases built from their articles. The lawsuit from the Murdoch-owned outlets also claims Perplexity attributed hallucinated quotes to real articles that never contained them. The New York Times issued a cease-and-desist, the BBC threatened legal action in June 2025, and three major Japanese publishers — Yomiuri Shimbun, Asahi Shimbun, and Nikkei — sued in August 2025, with Yomiuri alleging Perplexity "free-rode" on 120,000 of its articles in just five months.

The partnership track tells a different story. Perplexity launched its Publishers Program in July 2024, after more than a month of plagiarism accusations — including Forbes finding a lightly attributed copy of its paywalled reporting inside Perplexity's Pages tool. The program offers a double-digit percentage, flat-rate revenue share calculated per article, targeting 30 publishers by the end of 2024. Fortune, Time, Entrepreneur, The Texas Tribune, Der Spiegel, and WordPress.com signed on, and more than a dozen publishers expressed interest within two hours of launch, according to CNBC's coverage of the program.

Notably, News Corp sits on both sides of the divide. The same company suing Perplexity struck a multi-year licensing deal with OpenAI in May 2024 — proof that publishers will sell access when terms feel fair. CEO Robert Thomson's framing of the difference is blunt: he accused Perplexity of perpetrating "an abuse of intellectual property that harms journalists, writers, publishers and News Corp."

Then there's the rhetoric. CEO Aravind Srinivas has offered shifting explanations on scraping — first denying that Perplexity ignores robots.txt, then blaming third-party crawlers, then declining to commit to stopping that scraping of Wired's content. He has defended the product as curating content "akin to journalistic practice" while conceding it has "rough edges." Critics can reasonably ask whether a revenue share addresses the underlying problem when the alleged copying happened before any deal existed, and when independent investigations by Wired and Cloudflare allege the company disguised its crawlers to bypass blocking standards.

• Suing publishers argue Perplexity competes for their audiences using their own copyrighted work — with no permission and no payment.
• Partnering publishers accept that answer engines are here and negotiate a share of the value instead.
• Perplexity's own conduct — covert crawling allegations plus shifting CEO explanations — undermines the partnership pitch.
• The pattern isn't unique to Perplexity: licensing startup TollBit notes other AI companies also circumvent blocking standards, though Perplexity has drawn the most scrutiny.

For readers judging whether the fixes work, the test is simple: does revenue sharing change how content is acquired, or just how it's paid for after the fact? Until crawling practices and attribution mechanics change verifiably, the lawsuits will keep coming — and the burden of proof sits with Perplexity. It's the same verification challenge anyone building with AI faces, which is why SkillzBench is building independent verification — commit-pinned scanning, security checks, and benchmarking — into its free agent-skills directory, so trust rests on evidence rather than assurances.

The Verification Lesson: Trusting AI Outputs You Didn't Build

Strip away the lawsuits and the boardroom feuds, and the Perplexity controversy leaves every AI user with one uncomfortable question: how much do you trust an answer you didn't build? When an AI engine attributes hallucinated quotes to real articles — as the Dow Jones lawsuit alleges — the damage isn't just to the publisher's reputation. It's to your decision-making.

The scale matters here. At least 46% of Americans now use AI tools for information seeking, yet researchers describe hallucinations as "structurally inevitable" — and potentially more persuasive than human misinformation because of their fluent, authoritative tone. A confident citation attached to the wrong content is worse than no citation at all.

The lesson generalizes beyond search engines. Any AI output — a research summary, a code snippet, an agent's recommendation — carries hidden dependencies: where the data came from, who created it, and whether it was ever independently checked. Perplexity's own story shows what happens when those questions go unasked. After a month of plagiarism accusations, including Forbes finding its paywalled reporting reproduced with minimal attribution, the company scrambled to build the attribution and revenue-sharing structures it arguably needed from day one.

For teams building with AI agents, the same discipline applies to the components you plug in. That's the thinking behind SkillzBench's Verify pillar, which is being built to give every agent skill a documented provenance chain:

• Pin every skill version to a specific commit SHA, so you always know exactly what code your agent is running
• Run security scans and license checks before a skill ever reaches your workflow
• Publish a trust score built from security, quality, license, maintenance and community signals — with benchmark Lift always shown separately, so money never touches ratings
• Track upstream changes and re-scan when a skill's source updates, because "verified once" isn't "verified forever"

None of this is a knock on any single company. Perplexity's challenges — attribution, provenance, verification — are structural features of the AI era, not bugs of one startup. The winners, both as AI builders and as AI users, will be the people who treat every machine-generated output as a claim to be checked, not a fact to be accepted.

SkillzBench is building exactly that infrastructure for agent skills, and everything — the directory, the benchmarks, the free MCP connector — will be free to use. Join the waitlist to get early access, and when the platform opens, browse and test free agent skills through the free MCP connector to see verified, benchmarked skills in action.