The Wikimedia Foundation argues that restricting AI training too heavily would stifle open-source innovation, but permitting 'blanket' fair use could destroy the economic ecosystem that produces the information AI relies on. By separating the input stage (model training) from the output stage (RAG or chatbot responses), they aim to protect both technological progress and the secondary sources that contribute to collective knowledge.