AI Copyright: Wikimedia's Middle Ground Plea

The Wikimedia Foundation has adopted a nuanced stance in significant AI copyright litigation. The organization is urging courts to legally safeguard large language models (LLMs) during their training phases, even when using copyrighted materials. However, it distinguishes this from AI chatbots that might directly regurgitate protected content to users without proper licensing.

Wikimedia's legal filing expresses concern that current positions in these lawsuits oversimplify the fair use doctrine for AI. A victory for publishers could restrict transformative uses of work, while a win for AI companies might encourage disregard for licensing terms. The foundation highlights the existential importance of these cases for Wikipedia, noting its data is crucial for AI training, yet AI systems can divert users and potentially violate copyright by not adhering to Creative Commons license terms when reproducing content.

To navigate this complex landscape, Wikimedia proposes a compromise. It argues that scraping websites for data and using content for LLM pre-training without explicit licenses should be permissible if the LLMs create something new. Conversely, if AI output directly substitutes for original works in response to user queries, thus supplanting reader engagement, then licensing requirements should apply. This would ensure AI companies respect licensing obligations, particularly Wikipedia's Creative Commons terms for any adaptations or remixes of its content.