Pew: One-Third of New Web Pages Show Signs of AI Writing — What It Means for the AI Tools You Use

Introduction: The Web Hits the One-Third Mark

On August 20, 2026, Pew Research Center’s Data Labs published the most credible answer yet to a question everyone keeps asking: how much of the internet is actually written by AI? The report, “How Much of the Internet Is Written With AI?”, analyzed nearly half a million English-language webpages from the Common Crawl archive — collected between January 2021 and July 2026 — and ran each page’s text through an AI detection model built by Pangram.

The headline finding comes in two parts. In a random sample of 10,000 webpages from July 2026, 10% showed significant signs of AI authorship. But that figure mixes decades of old content that could never have been AI-written. Filter for pages published after ChatGPT’s November 2022 launch, and the share jumps to more than one in three. Less than four years after ChatGPT opened to the public — with roughly half of U.S. adults now using AI chatbots and 24% using them daily — AI has quietly become a co-author of a third of the new web.

The Numbers: What Pew Actually Found

Pew’s trend line tells a clean story: AI-authorship signals were flat until late 2022, then climbed steadily as ChatGPT, Claude, and Gemini arrived and multiplied. The domain breakdown is even more revealing:

When ChatGPT first launched, these linguistic signals appeared at similar rates across all major top-level domains. By 2026 they had diverged sharply — a fingerprint of where AI writing tools actually get deployed.

How Do You Detect AI Writing?

Pew used an open-weight detection model from Pangram that looks for statistical patterns in language — words, phrases, and linguistic quirks that AI models use more often than human writers. Some of the “tells” are remarkably specific: AI text overuses the em dash, loves to list items in threes, and reaches for the Oxford comma more than people do. Comparing today’s web with a 2023 snapshot, these tells have measurably spread across millions of pages.

The researchers are careful about limits: detectors misclassify individual documents in both directions, so no single page can be branded “AI-written” with certainty. But at the scale of hundreds of thousands of documents, the aggregate patterns are statistically robust — which is exactly how Pew framed its conclusions.

Why .com Domains Lead the AI Boom

The .com web runs on content volume: SEO-driven blogs, affiliate reviews, product pages, and marketing copy — precisely the workloads that AI writing tools are built for. It’s no coincidence that tools like ChatGPT, Jasper, Copy.ai, Writesonic, and Sudowrite dominate the commercial content stack. When a blog post costs pennies instead of an hour, the economic pull toward publishing more — and editing less — is enormous.

Meanwhile, .edu and .gov pages still show only ~1% AI signals — institutions with review processes and slower publishing cadences have absorbed AI very differently. The gap isn’t about access; it’s about incentives.

What This Means for Content Creators and SEO

A one-third AI share is the statistical backdrop to the “AI slop” debate that has consumed publishing all year — the same week, Google rolled out new ways for publishers to fight AI-driven traffic losses. When generic, well-formed, em-dash-heavy prose becomes the default, it also becomes worthless. Search engines are rewarding the inverse: first-hand experience, original data, and demonstrable expertise — the things a language model can’t fabricate.

The winning workflow in 2026 isn’t “human or AI” — it’s AI-assisted drafting with human expertise on top. If you’re building that stack, our guide to the best AI writing tools compares the leaders, and editing layers like Grammarly and QuillBot can strip the robotic tells before anything ships.

How to Use AI Writing Tools Without Shipping Slop

Pew’s data doesn’t say AI writing is bad — it says AI writing is everywhere, which makes differentiation the whole game. A practical checklist:

• Draft with AI, finish with expertise. Use AI for structure and first passes, then add first-hand knowledge, data, or reporting no model could generate.
• Fact-check every claim. Fluent text is not verified text — hallucinated statistics read exactly like real ones.
• Edit for voice. Cut the tells: the em-dash cascade, the rule-of-three lists, the hedge-everything tone. Sound like a person.
• Publish less, better. One genuinely useful page beats fifty thin ones — for rankings and for readers.
• Disclose where it matters. Trust is the scarce asset when a third of the web is machine-assisted.

Frequently Asked Questions