Wikimedia Draws a Line Between AI Training and Copyrighted Reproduction

Introduction

The central question in generative AI copyright disputes is shifting from whether companies may scrape websites to whether training and output should be judged in the same way. The Wikimedia Foundation has filed an amicus brief in a group of cases involving The New York Times, OpenAI, and Microsoft. Rather than backing either the publishers or the AI companies, it proposes a sharper distinction: learning from copyrighted works should not automatically be treated as infringement, while delivering the original work back to users should not be hidden behind the label of model training.

Key points

• Training and reproduction should be evaluated separately. Wikimedia argues that scraping and using copyrighted works for pretraining may qualify as fair use when the resulting model provides a new function or use that differs from the source material.
• Substitution creates a stronger copyright concern. If a chatbot retrieves article content in response to a query and gives users the same information without requiring them to visit the original page, the output may be replacing the source rather than transforming it. In that situation, licensing obligations should become more relevant.
• Open licensing still carries conditions. Wikipedia content is generally released under Creative Commons terms that include attribution and requirements for certain adaptations. Wikimedia says some AI services reproduce freely licensed text without following those conditions.
• Wikipedia is both a data source and an AI user. Large language models can help editors identify vandalism, possible copyright violations, doxxing, and articles that need improvement. At the same time, AI answers can divert readers from the sites and communities that sustain Wikipedia.

Why the distinction matters

Wikimedia fears that a sweeping publisher victory could allow copyright owners to gatekeep innovative and transformative uses of their work, making it harder for open-source tools and public-knowledge projects to operate. The opposite outcome carries a different risk: an overly broad victory for AI companies could be interpreted as permission to disregard licensing terms and weaken the bargaining position of authors and publishers.

The foundation’s proposal is therefore not a blanket declaration that training is always lawful. It asks courts to examine the practical result. Does the model provide a genuinely new use? Does the response replace the original work? Does the service preserve attribution and comply with the source license? Those questions may be more useful than a simple division between “AI innovation” and “copyright infringement.”

That framework is particularly important for Wikipedia. The encyclopedia depends on open, accessible human knowledge, but it also depends on readers visiting pages, contributing edits, and supporting the volunteer communities behind them. A legal rule that protects only one side of this relationship could weaken the entire information ecosystem.

The brief will not determine the outcome of the lawsuits, but it offers a possible direction for future policy. AI copyright rules may need to account for the training process, the generated response, source attribution, licensing, and whether an output diverts engagement from the original work. A more granular standard could protect meaningful model development without turning openly available knowledge into a free pass for commercial substitution.

Source: The Verge AI