Study: A Third of New Web Pages Were Written by AI Since the Launch of ChatGPT

The internet is no longer a space written solely by humans. With the widespread adoption of generative AI tools, a considerable proportion of web content has begun to be produced automatically without direct human intervention. In this context, a recent study issued by the Pew Research Center has put concrete numbers to a phenomenon that has long been discussed only in estimates.

The study concluded that more than a third of web pages published after the launch of the ChatGPT chatbot bear clear indicators that they were written or substantially edited by artificial intelligence, reinforcing the findings of earlier research that tracked the accelerating expansion of automatically generated content.

How Was the Study Conducted?

The researchers relied on the "Common Crawl" archive to collect nearly half a million English-language pages over a period of about five years, beginning roughly two years before the launch of ChatGPT in November 2022. The center then used a technology from "Open Pangram," a company specializing in detecting machine-generated text, with the aim of estimating the volume of content that was significantly written or edited using artificial intelligence.

Within a random sample of ten thousand pages collected in July 2026, about 10% of them showed clear signs of machine authorship. However, the researchers explained that this random sample inevitably includes older pages published before the emergence of AI writing tools, making it impossible for their source to be machine-based.

The Real Jump After Excluding Older Pages

To capture a more accurate picture, the center filtered out the older pages and focused exclusively on those published after the launch of ChatGPT. Here the result changed radically, as the proportion of pages bearing the fingerprints of machine authorship rose to more than a third, reaching 35%.

This figure reflects a structural shift in the nature of recently published content, and it intersects with a report issued by "Cloudflare," the internet infrastructure services company, which noted that bot traffic had surpassed human user traffic on the web—earlier than the company had expected. Thus we find ourselves facing a scene in which bots read pages written by other bots.

A Striking Disparity Between Domain Types

The proportions of machine-generated content were not equal across the various web domains; rather, the study revealed clear differences depending on the type of domain:

• .com domains recorded machine authorship indicators at a rate approaching ten times what appeared in educational and government domains.
• .edu and .gov domains were the lowest, at roughly 1% each.
• .org domains fell in the middle, at approximately 4.6%.

This disparity indicates that commercial websites rely on machine-generated content far more heavily than academic and government institutions, which are often subject to stricter editorial controls.

Linguistic Indicators That Reveal the Machine's Hand

In addition to the detection tools, the study observed an increase in certain linguistic traits that are often considered signs of machine writing, most notably the excessive use of em dashes, Oxford commas, as well as repetitive rhetorical formulas of the type "it's not this, but rather that."

The Study's Limits and Reliability

The center acknowledges that this analysis is not perfect, as machine-text detection tools—including "Pangram"—may sometimes mistakenly classify human-written pages as machine-generated. Nevertheless, the large volume of data makes the findings valid at least in terms of the general trend, and confirms that the web is undergoing a profound reshaping of its content sources.

These findings raise fundamental questions about the future of information quality and credibility on the internet, especially with the growth of a cycle in which artificial intelligence produces content consumed by other AI systems, which may affect the training of future models and the reliability of what reaches the end user.

✦ بقلم فريق دروب أيديا