Wikimedia urges court to separate AI training from outputs in copyright case
The Wikimedia Foundation has urged a US court to distinguish between AI training and AI outputs in copyright cases involving publishers, OpenAI and Microsoft. It says a broad ruling either way could harm access to knowledge: restricting training could limit useful tools, while allowing AI systems to reproduce protected material could weaken the open web and its creators.
In an amicus brief, Wikimedia argues that scraping and using copyrighted works to pre-train large language models can qualify as fair use when the models produce something new. But AI companies should need licences when their systems return material that substitutes for the original and draws readers away. Wikimedia says models can help Wikipedia editors, while also raising concerns about uncredited reuse of its Creative Commons content and reduced visits to its sites.
In short
- Wikimedia proposes different copyright rules for AI training and outputs.
- It says AI reproduction that replaces original works should require a licence.
- Wikipedia relies on open knowledge and says AI brings both risks and useful tools.
Steelman: both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Allowing AI developers to train on copyrighted works can support tools that help people find, summarise and create information, while broad restrictions could impede research and access to knowledge. If training is transformative and outputs do not substitute for the originals, treating it as fair use may preserve these benefits without requiring permission for every work used.
The case against
When an AI output reproduces protected material in a way that substitutes for the original, it can divert readers and revenue from publishers and creators. Requiring licences for such outputs could protect their ability to control and earn from their work, while addressing concerns about uncredited reuse and the sustainability of open websites.
Explained simply
AI companies train large language models on large collections of text, which can include copyrighted books, news and websites. The models use patterns in that material to generate answers, summaries and other text; those outputs may sometimes resemble or reproduce their sources.
The Wikimedia Foundation runs Wikipedia and other free knowledge projects. Its content is largely shared under Creative Commons licences, which set conditions for reuse, and the foundation says AI tools could help editors while also raising concerns about attribution and visitors being directed away from the original sites.
In US copyright law, fair use can allow some uses of protected work without permission, depending on the circumstances. The case raises questions about how those rules apply to training AI systems and to the material they produce, with implications for publishers, creators, technology companies and access to online information.
Read the full article at the source →
Originally published by The Verge as “Wikipedia’s operator says a bad AI copyright ruling could threaten the open web”.