Study Finds A Third of New Websites are AI-Generated

The more consequential detail buried in the finding is not the volume of AI-generated sites but the sentiment distortion: if aggregate web data is drifting toward artificially positive language, every model trained on crawled web data going forward is ingesting a subtly warped signal, compounding over each training cycle.

This connects obliquely to the Claude Mythos Preview story covered the same day from IEEE Spectrum. That piece showed a frontier model surfacing thousands of critical vulnerabilities without explicit security training, which raised questions about what capabilities emerge from large-scale exposure to production code and web content. If a third of new web content is now synthetic, the training surface for future models is increasingly self-referential: models trained on AI-generated text will produce models that generate more of it. The Claude Mythos finding is a reminder that emergent behaviors from web-scale training can surprise even the developers, and a systematically skewed web corpus is exactly the kind of upstream variable that makes those surprises harder to anticipate or explain.

Watch whether major model developers (Anthropic, Google, OpenAI) begin publishing explicit policies on filtering synthetic web content from training crawls within the next two quarters. A concrete filtering disclosure would confirm the industry treats corpus skew as a real risk rather than a theoretical one.