More Articles Are Now Created by AI Than Humans (graphite.io)

A new study using a 65k-sample from CommonCrawl finds that, as of November 2024, more articles published on the web are classified as AI-generated than human-written — a shift driven by rapid uptake after ChatGPT’s launch in November 2022. AI content growth climbed quickly (reaching roughly 39% within a year of ChatGPT) but has largely plateaued since May 2024, a trend the authors tentatively attribute to poor search performance for AI-written pages. They also note separately that AI-generated pages appear infrequently in Google and ChatGPT results, suggesting prevalence on the web may not translate to visibility or reader engagement. Methodologically, the study filtered English pages (article schema, ≥100 words, Jan 2020–May 2025) and used Surfer’s AI detector on 500-word chunks, labeling an article AI-generated if over 50% of chunks flagged as AI. Detection calibration found a 4.2% false-positive rate using pre-ChatGPT (Jan 2020–Nov 2022) human-era articles and a 0.6% false-negative rate on 6,009 GPT-4o–generated articles. Important caveats: the analysis doesn’t measure AI-assisted human-edited content (likely more common), evaluates false negatives only on GPT-4o, and warns that improving models could erode detector accuracy. The findings matter for ML practitioners because they affect the makeup of web training corpora, content-moderation and detection arms races, SEO strategies, and assumptions about real-user reach of AI-produced material.