The explosive growth of artificial intelligence-generated content that followed ChatGPT's launch has stalled, according to new analysis from digital marketing agency Graphite. The findings suggest that fears of AI-generated text overwhelming the internet may have been premature.
Graphite sampled more than 55,000 English-language articles and blog posts from Common Crawl, a major web archive, and ran them through three AI-detection tools. The results show that AI-generated articles surged rapidly after ChatGPT debuted in November 2022, climbing from roughly 36 percent of new online articles within the first year to 48 percent within two years. But since early 2025, that share has plateaued at around 50 percent.
The plateau marks a notable turning point in the AI narrative. Rather than progressively consuming more of the internet's written output, machine-written content appears to have hit a ceiling, at least for now. Human writers continue to produce roughly half of all new articles analyzed.
The implications concern researchers who study how AI models learn. Dan Klein, a UC Berkeley professor and AI model CTO, warned that training AI systems on machine-generated text creates a dangerous cycle. "These models are smart because of all the information we put on the web that was created without these models," Klein told Axios. "If we stop creating knowledge that is independent of these models, what's going to fuel that?"
The risk lies in what happens when AI trains on AI-generated content rather than original human knowledge. The result could be increasingly degraded output as models learn from their own recycled material rather than fresh human insights.
Measuring the exact share of AI-written content remains inexact. The line between human and machine authorship has blurred considerably. Many articles now involve human writers using AI for outlining, drafting, rewriting, or editing. Graphite classifies articles as primarily AI-generated only when most of the text is detected as AI-written or AI-assisted, but this standard doesn't capture the full spectrum of human-AI collaboration.
The quality of AI-generated content has also improved markedly. Graphite's analysis notes that machine-written articles often match or exceed the quality of human-written pieces, making them difficult for readers to distinguish. Despite these advances, the slowdown in AI article growth suggests factors beyond quality are at play, whether market saturation, reader preferences, or simply the practical limits of automating written content production at scale.
Author James Rodriguez: "The real story isn't that AI stopped growing, it's that the internet still needs humans to feed it ideas, and we're not producing them any faster."
Comments