Stanford: 35% of new websites created by AI - ForkLog: cryptocurrencies, AI, singularity, future

# Stanford: 35% of new websites created by AI

By mid-2025, about 35% of new websites were created entirely or partially with the help of artificial intelligence. This conclusion was reached by researchers at Stanford University.

Before OpenAI’s public launch of ChatGPT in November 2022, the indicator was at zero. Over a few years, the share of AI-generated content grew to more than a third of the latest publications on the internet.

The share of websites fully generated by AI (red), as well as created with neural networks (purple). Source: GitHub Researchers studied 33 months of archived website copies from the Wayback Machine using the Pangram v3 detector. The goal was to find out how the growth of AI texts is reshaping the structure of the World Wide Web.

Main changes

Researchers recorded a decrease in semantic diversity. Neural network-generated pages are 33% more similar to each other than texts written by humans. Different websites increasingly retell the same ideas with almost identical phrases.

According to the authors, it’s not just about mass AI copywriting. The problem is deeper: the variety of formulations and ideas is gradually narrowing. Large language models (LLM) by their nature select the most “average” responses and as a result reproduce template discourse.

The emotional tone of publications has also changed. AI content turned out to be 107% more positive than human content. At Stanford, this was linked to the already documented tendency of LLMs to flatter.

During training, developers optimize neural networks for pleasant, safe, and socially acceptable responses. As a result, a significant portion of new websites creates a “sterile friendly” information environment. There are fewer sharp judgments and conflicts, but also less lively human debate.

What was not confirmed

Several popular concerns did not find statistical confirmation. Researchers did not find a significant correlation between the growth of AI content and a decrease in factual accuracy, an increase in obvious errors, or stylistic homogenization of texts into a single template.

On the left: correlation between AI content and hypotheses. On the right: the share of adult Americans agreeing with each hypothesis. Source: GitHub. Researchers also pointed out a phenomenon that has so far been discussed mainly theoretically — model collapse (model collapse).

If new neural networks are trained on data with a lot of AI content, the system begins to digest its own averaged responses. This reduces variability, degrades quality, and threatens that in the future, LLMs will learn not from humans but from the “synthetic echo” of their predecessors.

Experts, together with the Internet Archive, plan to turn this research into a system for continuous monitoring of the share of AI content on the internet.

Recall that in mid-April, Stanford University pointed out the rapid pace of AI development. Researchers reported that neural networks have almost caught up with humans in performing computer tasks.