"Graphite used an AI detector called Surfer to analyze a random sample of URLs from Common Crawl, an open source database of over 300 billion web pages. The database spans 18 years and adds 3-5 billion new pages monthly. The pages had publish dates between January 2020 and May 2025 and were classified as either articles or listicles using Graphite's article page type classifier."
"Zoom in: Distinguishing between machine and human-written content is tricky. To evaluate Surfer's accuracy, Graphite tested it with its own sample of AI-generated articles and with a set published before ChatGPT's launch, which were likely written by humans. Surfer had a 4.2% false positive rate (labeling human-written articles as AI-generated) and a 0.6% false negative rate (labeling AI-written articles as human) for articles it generated with GPT-4o."
Graphite analyzed 65,000 URLs posted between January 2020 and May 2025 from Common Crawl to measure AI-generated content trends. Pages were classified as articles or listicles and labeled AI-generated when Surfer detected 50% or less human-written content. AI-generated article share rose sharply after ChatGPT's 2023 launch and briefly exceeded human-written articles in November 2024 before stabilizing near parity. Surfer's validation showed a 4.2% false positive rate and 0.6% false negative rate on Graphite's tests with GPT-4o samples and pre-ChatGPT human articles. Google Search and chatbot citations remained predominantly human-written, with 86% and 82% human authorship respectively.
Read at Axios
Unable to calculate read time
Collection
[
|
...
]