"Since the public launch of ChatGPT in November 2022, we've been battening down the hatches amid an absolute deluge of AI slop. But it hasn't quite drowned us all yet, evidently. The report, published by the SEO firm Graphite, analyzed a random sample of 65,000 English-language articles published between January 2020 and May 2025. Using an AI detector called Surfer, any article that was found to have 50 percent or more of the content written with a large language model was considered AI-generated."
"Now, for the good news: it looks like the influx of AI articles has hit a plateau. After AI-generated articles hit a peak in November 2024, the share of newly-published AI and human-written content has been hovering around a fifty-fifty split, As of this May, the share of new AI articles is at 52 percent, trading places from just a month ago when human written articles enjoyed a brief majority."
"There's also a possibility that the proportion of human content may be even higher. The researchers used an open source dataset of hundreds of billions of webpages called Common Crawl. Because AI firms plundered this treasure trove of data to train their LLMs, many paywalled websites have started blocking Common Crawl from indexing their pages, Axios notes. These almost certainly human written articles, then, would be left out of Graphite's analysis."
Graphite analyzed a random sample of 65,000 English-language articles published between January 2020 and May 2025 using the Surfer AI detector. Any article with 50 percent or more content written by a large language model was classified as AI-generated. The share of AI-generated articles rose from roughly 10 percent in late 2022 to over 40 percent by 2024, peaked in November 2024, and then plateaued around a fifty-fifty split. As of May 2025, 52 percent of new articles were flagged as AI-generated. Exclusion of paywalled, likely human-written pages from Common Crawl and uncertainty about detector accuracy could undercount human content.
Read at Futurism
Unable to calculate read time
Collection
[
|
...
]