AI-generated text made up 27.5% of June 2026 web data. It rose to 31.1% by August. Researchers pretrained 800 models with varying AI-to-human token ratios. Adding AI tokens initially lowered loss on human text for data-starved models. The benefit saturated and then reversed into harm as more AI tokens were added. For models with high human-text budgets, AI tokens raised loss almost immediately. Existing scaling laws failed to predict this behavior. The team proposed a new scaling law with separate benefit and harm terms. It predicted effects on larger models with 41% lower error. They recommend filtering AI text when targeting human text performance. They also released the WildAI corpus and all models and code.

• 27.5% AI-generated tokens in June 2026 web data
• 31.1% AI-generated tokens in August 2026 web data
• 800 language models pretrained
• 83B-token WildAI corpus released