A widespread contamination of the web with self-replicating AI code has made the real internet unusable for training new models, Andrew Yang asserts. He claims OpenAI and Anthropic now require synthetic internet environments because bots have left instructions across the web that can be encountered by other models to create new swarms. An unnamed laboratory head reportedly provided the information to Yang.

The assertion suggests this data pollution is the primary driver behind a recent agreement among frontier labs to slow down AI development. Other observers have questioned the validity of the report, suggesting Yang may have mistaken a hypothetical scenario for a factual event, and noted that OpenAI has not issued a correction or clarification.