Stanford Study: 35% of New Websites Created by AI by Mid-2025

Over a Third of New Websites Now AI-Generated

Researchers at Stanford University have determined that roughly 35% of new websites were created entirely or substantially with the help of artificial intelligence by mid-2025. Before the public launch of OpenAI's ChatGPT in November 2022, the figure hovered near zero. In just under three years, AI-generated content has grown to account for more than a third of fresh web publications.

The team analyzed 33 months of archived web snapshots from the Wayback Machine using the Pangram v3 detector. Their goal was to assess how the surge in AI-written text is reshaping the structure of the World Wide Web.

Why This Matters

The scale of AI content penetration directly impacts the entire information ecosystem. When more than a third of new publications originate from algorithms, the nature of knowledge and opinion that users encounter fundamentally shifts. For the crypto market and broader tech industry, this carries particular weight — a growing share of project analyses, token reviews, and market commentary may already be machine-generated, raising questions about the quality of information driving investment decisions.

Content Is Becoming More Homogeneous and Cheerful

One of the study's central findings is a measurable decline in semantic diversity. AI-generated pages are 33% more similar to one another compared to human-written content. Different websites increasingly restate the same ideas in nearly identical phrasing.

According to the authors, the root cause lies in the architecture of large language models (LLMs) themselves. These systems inherently gravitate toward statistically probable — and therefore "averaged" — responses. The result is templated discourse where the range of unique formulations and unconventional ideas steadily narrows.

The study also documented a shift in emotional tone. AI-generated content proved 107% more positive than human writing. The Stanford team attributed this to a well-documented tendency of LLMs toward sycophancy: during training, developers optimize models to produce pleasant, safe, and socially approved outputs. As a consequence, a significant portion of new websites creates a "sterile-friendly" information environment — one with fewer sharp opinions and genuine debate.

Which Fears Were Not Confirmed

Several widespread concerns failed to find statistical support. The researchers found no significant correlation between the growth of AI content and:

• a decline in factual accuracy;
• an increase in outright errors;
• stylistic flattening of all content into a single template.

The Model Collapse Threat

The researchers specifically highlighted the effect known as model collapse — a phenomenon that until recently remained largely theoretical. The mechanism works as follows: when new generations of neural networks are trained on data saturated with AI content, they begin recycling their own averaged outputs. This erodes variability and quality, and threatens a future where LLMs learn not from humans but from the "synthetic echo" of their predecessors.

The research team, in collaboration with the Internet Archive, plans to convert the study into a continuous monitoring system tracking the share of AI-generated content across the web.

Earlier in mid-April, the same Stanford group noted that AI development was outpacing expectations, with neural networks nearly matching human performance on computer-based tasks.