Wikipedia Takes a Middle Stance in AI Copyright Cases: Training Is Permitted, Verbatim Regurgitation Is Not

The Wikimedia Foundation, the organization behind Wikipedia, has not picked sides in the ongoing AI copyright lawsuits. It put forward a middle-ground proposal: the use of copyrighted works for large language model training deserves legal protection, but chatbots reproducing full original works directly for users should not enjoy the same protection.

Image credit: The Verge

As reported by The Verge, the Wikimedia Foundation submitted its amicus curiae brief on Friday to intervene in a set of landmark AI copyright cases. The parties on one side of these disputes include publishers such as The New York Times, while the other side consists of OpenAI and Microsoft. Wikimedia did not back either party, as it worries that the positions pushed by both sides will lead courts to issue excessively broad rulings.

Its position can be broken down into two simple points. Using copyrighted works to train LLMs should receive legal protection. Chatbots spitting out full copyrighted works verbatim to users should not be protected by the same set of legal logic. The former concerns whether Wikipedians can continue to use relevant AI tools, while the latter concerns whether copyright holders' works will be copied improperly through the output end of generative AI systems.

This distinction is very practical. All Wikipedia content is released by volunteers under open licenses, so both the open web and accessible AI tools are critical to the platform. If the court makes a blanket ruling that all training counts as infringement, Wikipedians may lose access to AI tools for editing, translation, and knowledge retrieval. If the court broadly exempts platforms from liability for infringing outputs, copyright holders will face the risk of their works being reproduced in full by chatbots.

Wikimedia stated in the brief that a black-and-white, all-or-nothing ruling could lead to one of two negative outcomes: either deprive Wikipedians of useful AI tools, or jeopardize the open web. It did not endorse OpenAI or the publishers, and only asked the court to evaluate training and output as two separate matters.

This middle path is not flashy, but it is a core issue that courts cannot avoid. Training and output are technically related, but they do not necessarily deserve the same legal ruling. The Verge's reporting lays out the core of this debate very clearly: AI copyright cases will not only determine how chatbots respond to users, but also affect whether open knowledge platforms can continue to access and use AI tools.