Wikimedia’s fair-use plan: license output that replaces the source

By Rebecca Stone

In a Friday amicus brief, the Wikimedia Foundation told the court considering The New York Times’s copyright lawsuit against OpenAI and Microsoft that pretraining large language models on copyrighted work can be fair use, while chatbot outputs that substitute for original articles require licenses.

That distinction defines Wikimedia’s middle ground in a group of landmark AI copyright cases: neither a publishers’ victory nor an AI victory would be safe. The foundation says both sides urge courts to apply fair use doctrine in overbroad, simplified ways.

A ruling for publishers could let copyright holders gatekeep innovative and transformative uses; a ruling for AI companies could encourage license avoidance, bypass negotiated terms and fragment the flow of knowledge between authors and audiences.

Wikimedia calls the stakes existential for Wikipedia, whose websites it describes as one of the most critical sources of data for all artificial intelligence and machine-learning training. The same systems draw readers away from those sites, weakening the contributor communities that maintain them, the foundation argues.

It also alleges that services such as ChatGPT reproduce Wikipedia text distributed under Creative Commons without complying with the license terms. The complaint cuts both ways: the nonprofit needs these models for useful work, including helping editors fight vandalism, detect copyright violations and doxxing, and identify articles that need improvement.

Wikimedia says a bad ruling could either deprive Wikipedians of useful AI tools or weaken the open web they sustain.

What separates fair-use training from licensed substitution?

Wikimedia proposes that scraping websites should not, by itself, constitute copyright infringement. It also says using copyrighted works to pretrain an LLM without a license should be fair use when the model produces something new and different from the original sources. The output supplies the test.

The threshold changes if a model pulls copies of protected material in response to a user query and delivers them while supplanting readers’ engagement with the original work. In that case, Wikimedia says the output substitutes for the source rather than transforming it enough to qualify as fair use, so a license should be required.

That obligation should fall on AI companies. The parties themselves remain divided.

The Times seeks a more expansive definition of fair use for LLMs and alleges that OpenAI infringed its copyright by scraping its website to build training datasets. OpenAI and Microsoft contend that LLM use is fair use. OpenAI has described output that regurgitates source work as a rare bug.

In a September filing, Microsoft said expert analysis found that Copilot rarely reproduces full sentences from news content used to ground the model. Microsoft argued that this finding hardly undermines the transformative purpose of LLM training.

What would licensing mean for Wikipedia?

For Wikipedia, licensing carries two concrete conditions. The Creative Commons terms require credit to the original source and require any remix or adaptation to remain under the same license. Wikimedia says AI companies that reproduce and redistribute material under those terms must follow both obligations, including when the material reaches users through a chatbot.

Wikimedia Foundation general council Stephen LaPorte put the wider risk this way: “A ruling that does not apply fair use principles carefully could threaten the sustainability of free knowledge, and stifle the development of open source technologies, ultimately impacting access to information for everyone in both cases.”

LaPorte added that Wikipedia’s collaborative creation depends on a strong foundation of open, accessible human knowledge.

Wikimedia CEO Bernadette Meehan said: “We are taking a position in this case to defend the free and open human knowledge that Wikipedia depends on.”

Expanding access through AI is powerful, she said, but it must respect the people, contributors and communities that make reliable knowledge possible. She described balancing those interests as a mutual benefit and shared responsibility across the information ecosystem.

The brief therefore does not ask for a blanket exemption for model builders or a blanket ban on training. Its compromise turns on the distinction between copyrighted input used to build a genuinely transformed model and output that replaces the original and competes for a reader’s attention.

The court’s treatment of that line will determine whether AI companies can continue training on openly licensed knowledge while facing licensing obligations when their products replace source works and draw readers away.