Creative Commons corpus ingestion into LLM training sets reached 2.1 billion licensed works by Q3 2024

The Chester Wisniewski post documents systematic scraping of CC-licensed images, text, and audio for commercial model training. Primary evidence appears in LAION-5B and Common Crawl snapshots where CC-BY and CC0 markers were retained yet downstream redistribution removed them. Model releases from Stability AI and Meta show no provenance logs for these subsets.

Data from the 2023 Common Crawl analysis and the 2024 EleutherAI data provenance study indicate 34 percent of sampled CC images carried no downstream license metadata after filtering. This pattern matches prior Common Crawl releases used for GPT-3 and Llama 2. Removal of attribution breaks the legal chain required by CC licenses that mandate notice preservation.

Operationally, creators who released work under CC terms now face direct competition from models outputting near-identical styles at zero marginal cost. Platforms hosting original CC content report measurable drops in referral traffic once synthetic versions appear in search results. No current model card discloses exact CC subset sizes or implements license-filtered inference.

Next measurable threshold is the outcome of ongoing EU AI Act conformity assessments due 2026, which require training data summaries. Absence of CC-specific filters in those filings will trigger enforcement actions against non-compliant providers.