**Algorithmic Law - Project/ You said: How much of the internet is taken up by AI writing, search the web?** How much of the internet is taken up by AI writing, search the web?Apr 20 **Claude responded: The short answer: it depends heavily on what you're measuring and where you look.**Synthesized multiple data sources to quantify AI content prevalenceSynthesized multiple data sources to quantify AI content prevalence The short answer: it depends heavily on what you're measuring and where you look. The numbers range wildly depending on methodology. The headline prediction vs. reality A 2022 Europol report circulated widely claiming up to 90 percent of online content could be synthetically generated by 2026 Futurism. That number became a viral talking point, but actual measurements haven't come close to it. What actual studies are finding The two most cited recent studies give different answers because they measure different things: • Graphite (SEO firm) analyzed 65,000 articles from the Common Crawl database and found that 50.3% of new web articles were primarily AI-generated as of November 2024, up from about 5% before ChatGPT eWEEK. By May 2025 the figure sat around 52%, and the growth has essentially plateaued. • Ahrefs analyzed 900,000 newly-crawled pages in April 2025 and found 74.2% contained AI-generated content Ahrefs. Note the key word — "contained," not "primarily written by." Hybrid human/AI content counts here. Where AI content actually shows up matters a lot This is the interesting part for your project. Even though roughly half of published new articles are AI-generated, they don't dominate what people actually read. Graphite's companion research found that among top-ranking pages in Google Search, only about 14% are AI-generated versus 86% human-written, and AI assistants like ChatGPT and Perplexity show a similar split at roughly 18% AI to 82% human. Caveats worth flagging AI detectors aren't perfectly reliable — they have false-positive rates (flagging human writing as AI) typically in the 4-5% range, and they struggle with lightly-edited AI output. So every number above is a rough estimate, not a measurement. The "50% of new articles" figure is probably the most defensible current claim, but it almost certainly undercounts lightly-edited hybrid content and overcounts formulaic human writing that happens to sound machine-like. For a project on signs of AI writing, the practical takeaway is that the volume problem is real but the visibility problem is smaller — search engines and chatbots are actively filtering toward human work, which is partly why recognizing the tells matters. You said: search the web for better sources, that are more well known? search the web for better sources, that are more well known?Apr 20 **Claude responded: Here's a cleaner picture drawing on more established sources — academic research, major news outlets, and the original EU report.**Curated reputable sources and organized key statisticsCurated reputable sources and organized key statistics Here's a cleaner picture drawing on more established sources — academic research, major news outlets, and the original EU report. The headline prediction everyone cites (and its limits) The "90% by 2026" figure that still gets quoted everywhere traces back to a 2022 report from Europol, the EU's law enforcement agency, which estimated that 90% of online content would be generated by AI by 2026 Axios. That was a speculative projection about deepfakes and synthetic media, not a measurement. Actual measurements have come in well below it. Axios, reporting on the Graphite study (the most cited current benchmark) Axios covered the most rigorous recent study, which looked at 65,000 URLs from the Common Crawl database between 2020 and 2025. Their findings: AI-generated articles briefly passed human-written ones in November 2024, and the two have remained roughly even since Axios. Importantly, Axios noted a methodological caveat — paywalled outlets like the New York Times block Common Crawl from indexing them, so human-written journalism is likely underrepresented in the sample. The University of Maryland / arXiv study (Oct 2025) — the most credible academic work A team from UMD, UMass, Microsoft, and Pangram Labs analyzed nearly 200,000 articles from 1,500+ US newspapers. The paper, "AI use in American newspapers is widespread, uneven, and rarely disclosed" arXiv, found: • Roughly 9% of news articles contained at least some AI-generated text • Major papers (circulation over 100,000) had much lower rates — only 1.7% — while smaller local outlets had significantly more • Opinion sections at the NYT, WSJ, and Washington Post were 6.4 times more likely to contain AI text than their newsrooms' reported journalism • Disclosure was nearly nonexistent — of 100 flagged articles they audited manually, only 5 acknowledged AI use Where AI content actually surfaces to readers The same Axios reporting noted that search engines and chatbots appear to be filtering it out: roughly 86% of top-ranking Google Search results and 82% of citations by ChatGPT and Perplexity trace back to human-written pages. The "mixed" problem (Slate, April 2026) Slate recently pointed out what I think is the most important nuance for a project like yours: in the Pangram newspaper study, 86.5 percent of the chatbot use detected in opinion pieces at the Times, the Wall Street Journal, and the Washington Post was classified as "mixed" Slate — meaning some unknown entanglement of human and machine. The binary "AI or human?" framing is already breaking down. Most "AI writing" in the wild now is hybrid, which is exactly what makes the stylistic tells in your quiz worth teaching. Cultural marker Worth noting: Merriam-Webster and the American Dialect Society both named slop their 2025 Word of the Year, which is roughly when the cultural awareness of this stuff crystallized. The most defensible summary If you want one sentence for your project: roughly half of newly published web articles show signs of AI generation, but only about 9% of actual US newspaper journalism does, and nearly all detected usage at major outlets is hybrid rather than fully machine-written — which is why recognizing the specific tells (not just flagging "AI vs. human") is the useful skill. Claude is AI and can make mistakes. Please double-check cited sources.