1/3 web pages published since ChatGPT's launch show signs of AI authorship ChatGPT发布以来三分之一的网页显示AI创作迹象

TL;DR

• Over one-third (35%) of web pages published after ChatGPT's November 2022 release show signs of AI authorship or substantial AI editing, according to a Pew Research study
• A random sample of 10,000 pages from July 2026 showed ~10% AI authorship overall, but this rose to 35% when pre-ChatGPT pages were filtered out
• .com domains exhibited AI authorship at roughly 10x the rate of .edu or .gov domains (~1% each), while .org domains sat at 4.6%
• The study used Common Crawl's web archive (nearly 500,000 English-language pages) and Open Pangram's AI-detection technology
• Secondary AI-writing indicators—em dashes, Oxford commas, and phrases like "it's not X, it's Y"—have also increased in prevalence over recent years

Why It Matters

This study provides one of the largest-scale empirical snapshots of AI-generated content on the open web, confirming that AI authorship has moved from a niche phenomenon to a dominant force in web publishing. For AI practitioners and researchers, it underscores the urgent need for robust detection, attribution, and content-authentication mechanisms as the internet becomes increasingly populated by machine-generated text.

Technical Details

• Data source: Nearly 500,000 English-language web pages collected from the Common Crawl web archive, spanning roughly five years beginning a couple of years before ChatGPT's November 2022 release
• Detection method: Open Pangram's AI-detection technology was used to classify pages as likely AI-written or substantially AI-edited; the study acknowledges that AI-detection tools can misclassify human-written content but argues the directional signal at scale is reliable
• Sampling methodology: A random sample of 10,000 pages from July 2026 yielded ~10% AI authorship; after filtering out pages published before ChatGPT, the rate jumped to 35%, demonstrating the importance of temporal normalization
• Domain-level analysis: .com domains showed ~10x the AI authorship rate of .edu and .gov domains (~1% each); .org domains registered 4.6%, suggesting commercial incentives drive the highest volume of AI-generated web content
• Linguistic markers: The study also tracked stylistic features associated with AI writing—em dashes, Oxford commas, and formulaic phrasing patterns—which have increased in frequency over the same period

Industry Insight

• The 35% AI authorship rate on post-ChatGPT web pages signals a fundamental shift in content ecosystems; platforms, search engines, and publishers should invest in content provenance standards (e.g., C2PA, watermarking) to maintain trust and differentiate human-created from machine-generated material
• The stark domain-level disparity (.com vs. .edu/.gov) suggests regulatory and academic spaces may remain relatively AI-clean for now, but commercial content farms and SEO-driven sites are prime candidates for AI saturation—organizations should audit their content supply chains accordingly
• As bot traffic has already overtaken human traffic (per Cloudflare), and a growing share of that content is AI-generated, the internet risks entering a feedback loop of bots reading bots; proactive investment in authentication, detection, and human-verified content channels will be critical to preserving information integrity

TL;DR

• ChatGPT发布后，35%的网页存在AI写作或大幅编辑迹象（排除2022年11月前页面）
• 整体网页样本中10%显示AI作者特征，.com域名AI内容比例是.edu/.gov的10倍
• 研究采用Common Crawl近50万英文网页数据，结合Open Pangram检测技术
• AI写作特征（如em dash、Oxford comma）随时间显著增加

为什么值得看

该研究量化了AI生成内容在网页中的渗透率，为内容创作者、平台运营者及政策制定者提供关键数据参考，揭示AI工具对网络内容生态的深远影响。

技术解析

研究基于Common Crawl网络档案，收集近50万英文网页（2017-2026年），使用Open Pangram技术检测AI写作痕迹。在2026年7月随机抽样10,000页中，10%显示显著AI作者特征；排除ChatGPT发布前的旧页面后，AI内容比例升至35%。域名分析显示.com站点AI内容占比约10%，而.edu/.gov仅1%，.org为4.6%。此外，AI写作典型语言特征（em dash、Oxford comma、"it's not X, it's Y"句式）的使用频率随时间显著上升。

行业启示

1. 内容平台需建立AI生成内容识别与标注机制，以维护信息可信度
2. 企业SEO策略应调整，避免过度依赖AI生成内容导致质量下降
3. 行业趋势表明AI写作工具已深度融入网络内容生产，需关注内容真实性与人类创作者价值重估