Language signals—and what they do not prove

• The model assessed combinations of linguistic patterns associated more often with AI than human text; Pew’s discussion includes increased em-dash and Oxford-comma use, AI-favored vocabulary, and negative parallelism as examples of such markers.
• Those are probabilistic signals, not authorship fingerprints. A human can use them, and AI-assisted text can be heavily edited; therefore the results should be read as “signs of AI authorship or meaningful AI editing,” not a definitive count of wholly AI-written pages.
• Detector error remains an important limitation: even tools with strong reported performance can produce false positives and false negatives, and external evaluations depend on the text types and models tested.

Scope limits

• Common Crawl covers publicly accessible pages, so paywalled and login-required sites are likely underrepresented.
• Only 10%–15% of sampled pages contained a usable publication-date field; the 35% figure therefore applies to that nonrandom dated subset.
• The study covered English-language web text, not images, video, non-English content, or the web as a whole.

Why it matters

Pew’s findings suggest that the web is becoming not merely a place where AI content exists, but one where recently published public text is increasingly AI-mediated. That shift coincides with Cloudflare’s reported crossover in which nonhuman traffic surpassed human traffic on its network in May 2026—although that traffic statistic concerns automated requests and bots, not AI-written content.

Together, the findings point to a web increasingly produced, crawled, and interacted with by automated systems—but they do not establish that AI has replaced human authorship or human web use overall.

Language signals—and what they do not prove

• The model assessed combinations of linguistic patterns associated more often with AI than human text; Pew’s discussion includes increased em-dash and Oxford-comma use, AI-favored vocabulary, and negative parallelism as examples of such markers.
• Those are probabilistic signals, not authorship fingerprints. A human can use them, and AI-assisted text can be heavily edited; therefore the results should be read as “signs of AI authorship or meaningful AI editing,” not a definitive count of wholly AI-written pages.
• Detector error remains an important limitation: even tools with strong reported performance can produce false positives and false negatives, and external evaluations depend on the text types and models tested.

Scope limits

• Common Crawl covers publicly accessible pages, so paywalled and login-required sites are likely underrepresented.
• Only 10%–15% of sampled pages contained a usable publication-date field; the 35% figure therefore applies to that nonrandom dated subset.
• The study covered English-language web text, not images, video, non-English content, or the web as a whole.

Why it matters

Pew’s findings suggest that the web is becoming not merely a place where AI content exists, but one where recently published public text is increasingly AI-mediated. That shift coincides with Cloudflare’s reported crossover in which nonhuman traffic surpassed human traffic on its network in May 2026—although that traffic statistic concerns automated requests and bots, not AI-written content.

Together, the findings point to a web increasingly produced, crawled, and interacted with by automated systems—but they do not establish that AI has replaced human authorship or human web use overall.