<figure class="wp-block-image size-large"><img width="325px" height="183px" srcSet="https://d2ly46a6b1u3g6.cloudfront.net/wp-content/uploads/2026/02/09114300/News-Human-Level-AI-Might-Arrive-in-Just-Two-Years_1-375x212.webp 375w, https://d2ly46a6b1u3g6.cloudfront.net/wp-content/uploads/2026/02/09114300/News-Human-Level-AI-Might-Arrive-in-Just-Two-Years_1-662x374.webp 662w, https://d2ly46a6b1u3g6.cloudfront.net/wp-content/uploads/2026/02/09114300/News-Human-Level-AI-Might-Arrive-in-Just-Two-Years_1.webp 686w" sizes="(max-width: 425px) 150px, 450px" src="https://ll-blog-dev.s3.ap-southeast-2.amazonaws.com/wp-content/uploads/2026/02/09114300/News-Human-Level-AI-Might-Arrive-in-Just-Two-Years_1-662x374.webp" alt="" class="wp-image-9220"/></figure>


A third of all new websites are AI-generated or AI-assisted by mid-2025. That&#x27;s according to a <a href="https://www.404media.co/study-finds-a-third-of-new-websites-are-ai-generated/" target="_blank" rel="noopener">new study</a> led by Stanford University and Imperial College London. The research analyzed Internet Archive data using the Pangram v3 AI detector. It found approximately 35% of newly published websites now rely on artificial intelligence. That&#x27;s up from essentially zero before ChatGPT&#x27;s debut in 2022.
The study reveals two significant changes reshaping the web.
First: semantic diversity is declining sharply. &quot;AI-generated pages are 33% more semantically similar to each other than human-written ones, meaning the same ideas and phrasings recur,&quot; according to the Stanford and Imperial College study. The same phrases. The same ideas. Over and over.
Second: AI-written text skews notably cheerful. &quot;AI text is markedly more upbeat, showing positive sentiment scores over 100% higher than human content,&quot; the researchers found. They attribute this to reinforcement learning. It rewards polite, agreeable responses.
Despite widespread concerns about misinformation, the data showed no significant rise in factual errors. AI content isn&#x27;t less accurate.
The study also debunked another popular worry. 83% of surveyed respondents feared AI would create a stylistic monoculture. Researchers found no evidence of character-level uniformity in writing styles. Not yet, anyway.
Interestingly, light or non-users of AI tools were more likely to believe in all hypothesized harms. Heavy users? Less worried.
More than a third of new sites are now AI-shaped. The researchers warn of &quot;model collapse&quot; risk. It&#x27;s a technical problem. Future AI models train on today&#x27;s increasingly AI-generated web. They learn from less diverse text. That could degrade their own performance.
The homogenization also carries broader cultural risks. An internet saturated with uniformly cheerful, semantically similar prose could subtly narrow public discourse. It crowds out dissent. No overt censorship needed.
The Stanford and Imperial College team plans to continue live monitoring with the Internet Archive. They&#x27;ll track AI&#x27;s growing footprint across the web.
AI tools are becoming more embedded in content creation workflows. Understanding their impact on information diversity matters. It matters for technical development. It matters for public policy.

Canton Network's Permission Controls Stir Decentralization Debate

Canton Network Addresses DeFi Security Concerns

EagleRock Land Targets $346 Million Raise In Permian-Focused IPO

Trump-Linked Crypto Project Sues Justin Sun for Defamation

AI-Generated Websites Now Account for One-Third of All New Sites

Stanford study reveals 35% of new websites use AI by 2025, raising concerns about semantic diversity and model collapse. Learn what this means for the web.