UC Berkeley 与华盛顿大学研究：截至 2026 年 8 月约 24% 可抓取网页含 AI 生成或编辑文本
New study from UC Berkeley and Univ of Washington found that about 24% of crawlable webpages held AI-generated or AI-edited text by August 2026, up from almost
UC Berkeley 与华盛顿大学的一项新研究显示，截至 2026 年 8 月，约 24% 的可抓取网页包含 AI 生成或 AI 编辑的文本，而 2022 年之前这一比例几乎为零。研究从 2019 至 2026 年的 Common Crawl 快照中抽取 510 万个各语言网页，用统计模型校正各检测器的误报与漏报，并以 2022 年前网页作为近乎无 AI 的基线。年增长几乎翻倍，从 2023 年初到 2025 年初的 4.6 个百分点升至此后的 8.1 个百分点；小型、少被链接的站点贡献了 78.8% 的增幅，而最核心的站点最不可能含 AI 文本。
正文
New study from UC Berkeley and Univ of Washington found that about 24% of crawlable webpages held AI-generated or AI-edited text by August 2026, up from almost none before 2022.
They sampled 5.1M pages from Common Crawl snapshots spanning 2019 to 2026, in every language. A statistical model corrected for each detector's false alarms and misses, using pre-2022 pages as a near-AI-free baseline.
Yearly growth nearly doubled, from 4.6 points between early 2023 and early 2025 to 8.1 points after that. Small, little-linked sites drove 78.8% of the rise, while the most central sites were least likely to have AI text.