Study: 'Wild' AI text breaks Chinchilla scaling laws

Researchers at UMass Amherst and Pangram Labs trained 800 language models and found that the value of AI-generated tokens flips sign: they help data-starved models but harm those already trained on large human corpora. By August 2026, 31% of web tokens were AI-generated.

• 31% of web tokens in August 2026 were AI-generated, up from 27.5% in June
• 800 language models trained with varying human-to-AI token ratios
• AI tokens raise loss once human data budget saturates
• WildAI corpus of 83B tokens released with Human/AI labels