AI Research Scientist

Pangram dreams of a slop-free future. Max Spero - a cofounder of Pangram, an AI detection tool popular among journalists - asks, "If you're not going to bother writing your newsletter, why should I read it?" In Do Androids Dream of Electric Sheep? - Philip K. Dick's dystopian 1968 novel, later adapted into the movie Blade Runner - a bounty hunter is tasked with killing cyborgs who look deceptively similar to humans. A polygraph-like device can distinguish between the two. Max Spero, the thirty-year-old cofounder and CEO of Pangram, an AI detection tool, likens his company to that machine. "In a sense," he told me, "we're detecting whether something is an android or a human." Spero - who calls himself "the internet's slop janitor" - met me at a Brooklyn café a few weeks after a big announcement for the company: Substack was partnering with Pangram to enable readers to scan posts and test whether the writing had been produced using generative AI. "When readers have to wonder if what they're reading is real, it undermines trust in authorship and threatens the livelihood of writers - including those who use AI tools thoughtfully to produce work they believe in," Chris Best, a cofounder of Substack, wrote in a post announcing the partnership. "Platforms that reward fakeness will create a race to the bottom." Spero met Best in June through Jasmine Sun, who writes about AI for The Atlantic and on her own Substack, which has just over thirty thousand subscribers. Sun ran into Best at a Substack event - where she gave a talk on independent writing in the age of AI - and told him that he should think about partnering with Pangram. "Basically every time I ran into someone from Substack, I'd be like, 'You guys should work with Pangram,'" Sun, who'd previously worked as a product manager at Substack, told me. (The night before the event, she had attended a dinner hosted by Pangram.) Best, who had been worried about slop, asked Sun to make the introduction to Spero. Central to Pangram's ethos is the proposition that the internet should be a more honest place. "I think there are cases where AI is used to synthesize massive amounts of data and it's a very useful tool for research; there are also cases where somebody hasn't bothered to write their newsletters themselves," Spero told me. "If you're not going to bother writing your newsletter, why should I read it?" Spero founded Pangram in 2023 alongside Bradley Emi, who is now the company's chief technology officer. (Previously, he'd worked at Google and Neuro, the self-driving-car startup.) Since then, Pangram has grown popular among writers, journalists, and academics. A 2025 study by the University of Chicago's Becker Friedman Institute for Economics found that Pangram "achieves essentially zero false positive rates and false negative rates on medium-length to long passages." In April, Taylor Lorenz, the technology journalist, called it the "best tool on the market." The Substack partnership is one of many recent advances for the company. This spring, Pangram announced a deal with NewsGuard, which rates outlets' biases; it also works with PubShield, a platform that scans manuscripts for AI use and plagiarism, and Tremau, a content moderation company, among others. In July, Menlo Ventures, a venture capital firm, raised nine million dollars for Pangram. Pangram's rise comes amid an outcry over AI-generated writing. (Claude, Anthropic's large language model, is now adding a watermark to AI-generated text.) But Pangram does have its detractors, some of whom believe that the tool is inaccurate, even dangerous: "Substack's newest subscription model: guilty until proven human," a content creator wrote. "What is Pangram actually for?" asked another. "Virtue signaling and witch hunts." Spero said countering that narrative has been challenging. People think, "Oh, there's been a dozen AI detectors that sucked. Why should we believe yours is good?" he told me. Lately, he added, some Substackers have said the tool "consumes a bunch of water" and "vacuums up" their data. He disputes these claims. "It's very easy," he said, "to make up lies on the internet." Spero is particularly interested in how Pangram can be used to strengthen journalism. He has friends who are writers, he told me, and knows it's a "rough industry where there's so much pressure to do more with less." In the past year, journalists writing for a range of publications, including the New York Times and Fortune, have admitted to using generative AI. Researchers at the University of Maryland, who used Pangram to analyze a hundred and eighty-six thousand articles published in fifteen hundred American newspapers in the summer of 2025, found that more than 9 percent contained AI text. The same study found AI-generated text in the opinion sections of the Times, the Washington Post, and the Wall Street Journal. As generative AI use in writing becomes more common, so does a need for transparency and accountability, Spero told me. Pangram, he hopes, will help meet that need - and inspire journalists and publications to create guidelines around the use of generative AI. "Hopefully," he said, "we're doing our own small part to keep journalism human." Editor's note: This article has been updated to correct the publication date of Do Androids Dream of Electric Sheep? Has America ever needed a media defender more than now? Help CJR by joining CJR today. Carolina Abbott Galvão is a Delacorte fellow at CJR.

35% of web pages published since ChatGPT show AI authorship signs. August 20, 2026 35% of web pages published since ChatGPT are now ai-written. A new study from Pew Research has quantified what many suspected: artificial intelligence is now writing a massive portion of the internet. Analyzing nearly half a million English-language web pages from the Common Crawl archive, Pew found that over one-third (35%) of web pages published after ChatGPT's November 2022 release show significant signs of AI authorship. For the broader web - including older pages that predate AI writing tools - the figure stands at roughly 10%. But when you isolate content published in the AI era, the number jumps dramatically. The internet is being reshaped in real time, and the data reveals just how fast. How Pew measured AI authorship. Pew Research partnered with Pangram, an AI-detection technology provider, to analyze a random sample of 10,000 web pages collected in July 2026. The methodology involved: * Pulling ~480,000 English-language web pages from Common Crawl, spanning approximately five years * Filtering to isolate pages published after November 2022 (ChatGPT's release) * Running Pangram's detection model to classify pages as likely AI-written, substantially AI-edited, or human-authored * Cross-referencing domain types (.com, .edu, .gov, .org) to identify where AI content concentrates Pew acknowledged that AI detection tools like Pangram can misclassify content - some human-written pages may be flagged as AI, and vice versa. However, at the scale of hundreds of thousands of pages, the data is directionally reliable. The trend is unmistakable. The domain divide: .com vs .edu. One of the most striking findings is how AI authorship distributes across domain types: * .com domains: ~10% AI-authored (the highest rate) * .org domains: 4.6% AI-authored * .edu domains: ~1% AI-authored * .gov domains: ~1% AI-authored Commercial websites are generating AI content at 10x the rate of educational and government sites. This gap reflects a fundamental economic incentive: .com domains are under constant pressure to publish more content for SEO, marketing, and customer acquisition. AI writing tools dramatically lower the cost and time barrier, making it rational for businesses to produce content at scale. Meanwhile, .edu and .gov domains - governed by editorial standards, peer review, and institutional oversight - have barely budged. The contrast reveals where AI content is driven by utility versus where it's driven by volume incentives. The linguistic fingerprints of AI. Beyond detection tooling, Pew identified specific linguistic patterns that have surged since ChatGPT's release: * Em dashes (-): Usage has increased significantly, as AI models favor this punctuation for parenthetical clauses * Oxford commas: Once declining in casual web writing, now resurgent - a hallmark of AI-generated prose * "It's not X, it's Y" constructions: A rhetorical pattern AI models deploy frequently, now appearing far more often in web content These aren't coincidences. Large language models trained on curated corpora tend to produce grammatically polished, stylistically consistent text. When millions of pages are generated or edited by these models, the statistical fingerprint becomes visible at the web scale. Bots reading content written by bots. The Pew study arrives alongside another milestone: Cloudflare reported that bot web traffic has overtaken human web traffic for the first time. This means the internet is increasingly a machine-to-machine environment - bots crawling pages that were themselves written by AI. The implications for web infrastructure are significant: * Server load from AI crawlers: Search engine bots, AI training crawlers, and content-scraping agents now generate the majority of HTTP requests. Servers must handle this traffic without degrading performance for human visitors. * Content inflation: As AI makes content production nearly free, the total volume of web pages is exploding. Hosting providers face increased storage and bandwidth demands. * Quality vs quantity tension: More content doesn't mean better content. The 35% AI-authored figure raises questions about content quality, originality, and the long-term value of the web as an information source. What this means for website operators. For businesses running websites, the AI content wave creates both opportunities and risks: Opportunities: * AI-assisted content creation can dramatically reduce the cost of maintaining a content-rich website * Small teams can now publish at a volume that previously required dedicated content departments * AI tools can help maintain consistency across large content libraries * Search engines are increasingly sophisticated at detecting low-value AI content. Google's helpful content updates have already penalized sites that publish generic AI-generated articles at scale. * AI content that is factually incorrect - known as hallucination - can damage brand credibility and mislead users. * The proliferation of AI content creates a "noise floor" that makes it harder for genuinely useful content to stand out. The data labeling economy behind the AI wave. The AI content explosion is fueled by a parallel boom in AI training data. Companies like Micro1 have seen explosive growth - Micro1's gross annual run rate jumped from $100 million to $500 million in just eight months, driven by demand from AI labs for high-quality training data. Competitors like Mercor ($2 billion gross revenue) and Handshake ($1 billion) show that the data-labeling market alone is now a multi-billion dollar industry. This means the AI writing tools generating 35% of new web pages are themselves being trained on increasingly large and sophisticated datasets - creating a feedback loop where AI content trains the next generation of AI, which produces more AI content. Implications for the future of web content. The Pew data suggests HostMop is at an inflection point. If 35% of post-ChatGPT web pages are AI-authored today, that percentage is likely to increase as tools become more accessible and businesses adopt them more aggressively. Several trends to watch: * AI content detection arms race: As AI writing improves, detection tools will struggle to keep up. Pew's reliance on Pangram highlights the cat-and-mouse dynamic between generation and detection. * Regulatory pressure: The EU has already ruled that AI-generated content is not protected by copyright. Other jurisdictions may follow with labeling requirements or restrictions on undisclosed AI content. * Infrastructure adaptation: Web hosting and CDN providers will need to optimize for a web where the majority of traffic and a growing share of content is machine-generated. This includes handling aggressive AI crawler behavior, managing content storage at scale, and ensuring human-facing performance isn't degraded. The bottom line. Pew Research's findings confirm what the web industry has felt anecdotally: AI is no longer a novelty in content creation - it's a core part of the web's content pipeline. With 35% of new pages showing AI authorship signals and bot traffic exceeding human traffic, the internet is becoming a fundamentally different medium. For website operators, hosting providers, and infrastructure teams, the message is clear: the web of 2026 is built on AI-generated content, crawled by AI-driven bots, and served through infrastructure that must adapt to this new reality. Understanding these shifts is essential for anyone whose business depends on the web - which, increasingly, is everyone. Topik terkait. AI content AI authorship Pew Research web content ChatGPT web infrastructure bot traffic Need hosting that can run an AI Agent 24/7? VPS + AI Agent from $39/month. Deploy in 5 minutes, instantly working.

Anthropic says it will watermark text generated by its AI models. Anthropic will extend support for watermarking AI generations for older models as well. Anthropic will watermark text generated by its models, including Claude, to comply with European regulations, the company now says. The AI model maker confirmed the watermarking in an updated support page. EU AI Act's Transparency Code, which took effect on August 2, requires AI companies to mark AI-generated or edited content in a way other systems can identify them. Anthropic said that all models released after August 2 will automatically have tech to watermark both computer-generated text and files. For files, the company is using the C2PA open standard. The company said it will extend support for older models as well, adding that the watermark will travel when users copy and paste the text. "Because the watermark is part of the text, it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing. Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from," the support page reads. It's not clear how much editing users need to do to remove the watermark. IntelPro has asked Anthropic to clarify and will update the story if IntelPro hear back. The company noted that watermarking will apply to different products like Claude platform API, Claude, Claude Code, Claude Cowork, and Claude Tag. Platforms are now rushing to watermark AI-generated content after backlash from users and to avoid regulatory scrutiny. Last week, AI music platform Suno said it will mark tracks created on its platform after a spate of legal challenges. Last month, newsletter service Substack teamed up with Pangram to flag AI-generated content. The company's CEO, Chris Best, called out Claudefishing, a term used for people using AI to generate content. Apart from Anthropic, other companies like Black Forest Labs, Google, Meta, Microsoft, OpenAI, and Synthesia have committed to adhering to the EU's code.