Wikimedia requests legal clarification on copyright in AI training vs. content reproduction in chatbots amid lawsuits against OpenAI.
On October 30, Wikimedia, the organization behind Wikipedia, filed a motion in a federal court in New York to clarify the legal boundaries surrounding the use of copyrighted materials for training artificial intelligence versus reproducing that content in chatbot responses.
This appeal comes amid a collective legal action involving twelve lawsuits against OpenAI, including one from The New York Times, which accuses OpenAI and Microsoft of unlawfully utilizing millions of its articles to train AI models and generate responses. While Wikimedia has not sided with either party in this dispute, it expressed concern that an overly broad interpretation of copyright law could hinder both AI development and the availability of open knowledge.
Key Legal Issues
The central issue at hand is whether copyrighted works can be used for training large language models without the permission of copyright holders. The New York Times initiated the lawsuit against OpenAI and Microsoft in December 2023, alleging copyright infringement due to the use of its articles.
The defendants argue that using publicly available materials for AI training is a transformative use and qualifies as fair use under copyright law.
Wikimedia proposed that the court evaluate two distinct scenarios:
- The preliminary training of language models using internet texts, where the use of copyrighted works could be deemed fair if the output significantly differs from the original and does not replace it.
- The generation of responses that incorporate extracted data, where AI utilizes information from external sources to prepare its replies.
According to Wikimedia, if an AI system extracts and reproduces copyrighted content in a way that substitutes the original work, such use should not automatically be classified as fair use, and developers must adhere to licensing requirements.
Wikimedia's Neutral Stance
Wikimedia warned of two opposing risks in its legal position. If the court rules that AI training on copyrighted works constitutes infringement, it could stifle the development of innovative AI models, including open-source ones. Conversely, if the court broadly allows the use of third-party content as fair use, it may diminish incentives to create original content.
This could enable AI developers to present information without directing users to the original sources, jeopardizing the sustainability of the open internet that relies on the efforts of journalists, researchers, and volunteer authors.
“A decision that fails to properly consider fair use principles could threaten the sustainability of free knowledge,” stated Wikimedia's legal counsel, Stephen Laporte.
Currently, Wikipedia hosts over 67 million articles across 300 languages, with approximately 250,000 volunteers contributing each month to its content.
Implications for Wikipedia
Wikipedia operates under the Creative Commons Attribution-ShareAlike 4.0 license, which permits the copying and modification of its content, including for commercial uses, provided that attribution is given and adaptations are shared alike.
Wikimedia argues that using its articles for AI training might fall under fair use without complying with licensing terms. However, if chatbot developers reproduce Wikipedia materials in their responses, they should follow the terms of the CC BY-SA 4.0 license, including attribution requirements.
Additionally, Wikimedia urged developers to consider the impact on Wikipedia's infrastructure when conducting automated data collection, advising adherence to public API limitations or the use of the Wikimedia Enterprise API.
The organization submitted its conclusion as an amicus curiae, which means its position is not legally binding and does not imply that AI training is recognized as fair use or that the use of retrieval-augmented generation (RAG) constitutes copyright infringement.
