Nearly half of sampled English-language articles published online were classified as mostly AI-generated by three commercial AI detectors, according to a new study that analyzed 55,000 webpages from Common Crawl.
- Graphite found that detector-classified AI-written content surged after ChatGPT launched in late 2022, then stabilized at roughly 50% of sampled articles through early 2026.
- Articles classified as mostly AI-generated accounted for an estimated 49.9% of sampled content in Q1 2026, the study found.
- In late 2025, detector-classified AI-generated articles briefly surpassed detector-classified human-written ones.
Is AI publishing slowing? Mainly AI-generated articles accounted for 35.9% of sampled published content within 12 months of ChatGPT’s November 2022 release, the study found. That figure rose to about 48% by late 2024, then stabilized near 50% starting in Q1 2025.
- The slowdown may suggest publishers are learning that heavily AI-generated articles don’t always perform well in search, Graphite said. However, the study didn’t directly measure search rankings, traffic, or visibility.
Why we care. The findings suggest AI-generated or AI-assisted publishing is now common across the web, especially in large-scale content workflows. However, the study doesn’t show whether AI-generated articles receive more or less traffic than human-written content, or whether people regularly encounter them in Google Search or AI systems.
What the study measured. Graphite analyzed 55,400 English-language articles from Common Crawl published between January 2020 and March 2026. Articles were classified using three AI detection systems:
- Pangram
- Copyleaks
- GPTZero
Graphite then averaged the results from all three systems. The study defined “primarily AI-generated” differently for each detector:
- Pangram and Copyleaks estimated how much of an article was AI-generated.
- GPTZero used article-level labels such as “Human,” “Mixed,” and “AI.”
- Articles labeled “Mixed” by GPTZero were counted as primarily AI-generated.
- All three detectors showed the same broad pattern: a sharp rise in AI-classified publishing after ChatGPT’s release, followed by a plateau near 50%.
Limits and validation of the AI detectors. To estimate false positives, Graphite tested the detectors against 15,700 articles published before ChatGPT launched in November 2022, assuming those articles were mostly human-written. To estimate false negatives, researchers generated 6,000 AI-written articles using GPT-5, Gemini 3.1 Pro, and Claude Opus 4.6.
- Average false-positive and false-negative rates were below 2% across all three detectors in those benchmark tests.
- However, the study didn’t evaluate mixed workflows in which humans heavily edit AI drafts — an increasingly common publishing process that may be harder for detectors to classify accurately.
- The researchers also said future AI models may become harder to detect as generation quality improves.
Search Engine Land is owned by Semrush. We remain committed to providing high-quality coverage of marketing topics. Unless otherwise noted, this page’s content was written by either an employee or a paid contractor of Semrush Inc.