Wikipedia AI Copyright has spent the past two years as a courtroom abstraction: who gets to own the right to train machines on human knowledge. This week the abstraction grew an institution. Wikimedia, the nonprofit that operates Wikipedia, stepped into the lawsuits as a friend of the court, warning that both the publishers and the AI labs are steering toward outcomes the open web would not survive.
Its proposed Wikipedia AI Copyright compromise splits the legal question in two — and asks every judge to draw one uncomfortable line.
- The Wikipedia AI Copyright Brief, Explained
- Why Wikipedia Sits in the Middle
- The Fine Line Wikimedia Wants Courts to Draw
- The Lawsuits That Could Rewire the Internet
- The Creative Commons Promise in the Crosshairs
- What a Publisher Victory Would Break
- What an AI Victory Would Break
- Why the Wikipedia AI Copyright Brief Matters to You
- References
The Wikipedia AI Copyright Brief, Explained
The nonprofit behind Wikipedia just did something rare in the Wikipedia AI Copyright fight over whether artificial intelligence may learn from human work: it refused to pick a side. In an amicus brief filed Friday in a cluster of landmark lawsuits pitting publishers including The New York Times against OpenAI and Microsoft, the Wikimedia Foundation laid out a Wikipedia AI copyright position that concedes fair use protects model training, while insisting that chatbots which recite copyrighted text back to users need licenses. One badly drawn line, it warns, could either strip Wikipedians of AI tools they depend on or hand AI companies a green light to gut the open web.
What the Brief Actually Says
At its core, the Wikipedia AI copyright filing splits one question into two. First: is scraping websites and building training datasets from the results protected fair use? Wikimedia answers yes, provided the model produces something genuinely new rather than a photocopy with grammar. Second: what happens at the output stage, when a model reproduces material in response to a prompt and delivers it to readers who would otherwise have visited the original source? There, the brief argues, a license should be required.
The distinction turns on substitution. A model that digests millions of articles and answers questions in ways the originals never did has transformed the material, the brief says. A model that hands a user an article-shaped chunk of text that satisfies the same appetite the original would have satisfied has simply replaced it, and replacement is where fair use tradition grows weak.
Why Neither Side Won This Framing
Wikimedia is explicit that it is not a party with a stake in who collects damages. It is a party with a stake in the rule itself. The brief states that the parties positions push for overbroad and simplified applications of the fair use doctrine, and warns that a clean victory for publishers could let copyright holders gatekeep even innovative and transformative uses of their own work. A clean victory for the AI companies, conversely, would be a green light to snub licensing terms and fragment the flow of knowledge between human authors and audiences.
| Party | Training on copyrighted work | Outputs that echo sources |
|---|---|---|
| The New York Times | Infringing scraping; demands expansive protection | Core evidence of copying |
| OpenAI and Microsoft | Fair use; transformative by definition | A rare bug, not a feature |
| Wikimedia Foundation | Fair use when the result is new and different | Licensed or attributed when it substitutes for the original |
That middle lane of the Wikipedia AI Copyright fight is uncomfortable, and Wikimedia walked into it anyway because the existence of the road, more than any verdict, decides what its project becomes.
Why Wikipedia Sits in the Middle
The Traffic Problem
The Wikipedia AI copyright argument begins with an uncomfortable symmetry: Wikipedia is simultaneously one of the internet’s greatest donors to AI training data and one of its most visible traffic casualties. The brief itself notes that Wikimedia websites have become one of the most critical sources of data for all artificial intelligence and machine learning training. Meanwhile, the chatbots that drank from that well are answering questions directly, and the readers who once clicked through to articles are not arriving.
For a commercial publisher, a vanished reader is lost revenue. For Wikipedia, it is something stranger: a vanished volunteer. Every article on the site exists because someone spent unpaid hours writing, citing and patrolling it. When AI systems absorb that labor and return answers without the attribution loop, they weaken the very feedback of readers who might otherwise become editors.
The Licensing Loophole
Wikimedia adds a second grievance that is easy to miss. Its text is licensed freely to anyone under Creative Commons terms, but freedom is conditional: users must credit the source and pass the same license along. Chatbot answers that reproduce Wikipedia text without credit are, the foundation argues, violating a license that was generous by design. The Wikipedia AI copyright brief frames this not as hostility to AI but as the opposite: a demand that AI companies honor the one permission structure that already lets them use its work for almost free.
Yet the foundation is equally blunt that the other half of its interest is real. The same models that siphon readers also help editors fight vandalism, spot copyright violations and doxxing, and find articles that need improvement. Ban the technology, and those guardians lose their tools first. That framing is the heart of its Wikipedia AI Copyright argument.
The Fine Line Wikimedia Wants Courts to Draw
Training That Transforms: Fair Use
The compromise Wikimedia proposes reads like a lawyer’s scalpels. Scraping websites, by itself, should not be copyright infringement, the Wikipedia AI copyright filing argues. Neither should unlicensed pre-training, so long as the resulting model offers something new and different from the original sources. That phrase matters: it borrows the transformation language courts have used for decades in parody and scholarship cases, and stretches it across machine learning at industrial scale.
Outputs That Substitute: License Required
The Wikipedia AI Copyright line snaps shut at delivery. When a model pulls a copy of material in response to a user query and hands it over while supplanting the reader’s engagement with the original work, the brief says that output should require a license. In those cases, AI companies should respect licensing requirements, and for Wikipedia content that means honoring Creative Commons: credit the source, and keep any remix under the same license.
| Test | Training stage | Output stage |
|---|---|---|
| What happens | Scraping, weighting patterns inside billions of parameters | Model returns text echoing a specific source |
| Does it replace the original? | No reader consumes the source instead | Yes, the reader stops going to the source |
| Wikimedia verdict | Fair use, if the result is new and different | License or attribution required |
General counsel Stephen LaPorte put the stakes plainly: 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. Wikimedia hopes courts will formalize that Wikipedia AI Copyright distinction.
The Lawsuits That Could Rewire the Internet
The Times Argues for a Wider Fence
The litigation that gives the Wikipedia AI copyright filing its stage is led by The New York Times, which sued OpenAI and Microsoft in late 2023, alleging that scraping its archives to build training datasets was straightforward infringement. The Times wants courts to read fair use narrowly for large language models, so that copying at that scale is the violation, full stop. It has kept adding evidence since, and our own coverage of the USA Today lawsuit against OpenAI shows how steadily the publisher coalition has widened.
OpenAI and Microsoft Fight Back on Fair Use
OpenAI and Microsoft answer that training is fair use because the models create new expression, and that regurgitating copyrighted text is, in OpenAI’s phrase, a rare bug rather than the product. Microsoft’s September filing points to expert analysis finding that Copilot rarely reproduces full sentences from grounding content, which it says hardly undermines the transformative purpose of LLM training. Both sides want the court to ratify their whole theory of the information economy.
What is getting lost between the Wikipedia AI Copyright briefs is the smaller web. Wikipedia’s traffic chart shows why its lawyers thought the moment urgent: the readers who fund and volunteer for free knowledge are leaving in numbers that no longer look temporary.
A ruling written only for newspapers and chatbot giants could freeze whatever number this curve reaches next, for every free publisher still standing.
One underappreciated wrinkle: the cases are spread across different federal circuits, each with its own precedent on fair use. A ruling in New York does not bind a court in California, so the same Wikipedia AI Copyright question could produce genuinely different answers depending on where a lawsuit lands. That patchwork outcome is precisely what Wikimedia warns is unworkable for global platforms, and it raises the odds that the fight over training data ends up in front of the Supreme Court regardless of who wins the first round.
The Creative Commons Promise in the Crosshairs
A License Built on Trust
Creative Commons is the quiet operating system of the open web: tens of millions of works shared on the promise that users will credit authors and keep derivatives free under the same terms. The Wikipedia AI copyright brief is, beneath the legal argument, a plea from the biggest beneficiary of that promise. Wikimedia text is free for anyone, even AI companies, so long as they follow conditions that cost nothing to observe. Chatbots that recite it without credit are, the foundation says, freeloaders in a system that only works when everyone pays the small toll.
The Wikipedia AI Copyright stakes reach past any single verdict. If courts bless uncredited reproduction, the moral contract behind the CC BY-SA license collapses into a suggestion. Every musician, photographer and educator who released work under those terms watching a model resell their output with a straight face will draw conclusions about the next upload.
Share-Alike Versus Synthetic Text
There is a second Wikipedia AI Copyright puzzle Wikimedia raises. Share-alike requires adaptations to carry the same license forward. A model that remixes encyclopedic text into answers arguably produces an adaptation, which would mean those outputs owe the same license in return. Applying that logic at planetary scale would be administratively absurd, which is exactly why the Wikipedia AI copyright filing prefers its simpler remedy: enforce attribution and licensing at the output boundary, where substitution happens, instead of trying to trace parameters back to sources.
If that sounds like a compromise designed to be technically imperfect and practically survivable, that is the point. The open web was built by imperfect people agreeing to simple rules.
What a Publisher Victory Would Break
Gatekeeping Transformative Work
Imagine the court adopts the publishers’ Wikipedia AI Copyright framing wholesale: scraping is infringement, and training is not transformative no matter what the model does afterward. Every researcher who builds a dataset without paying license fees becomes a defendant. Every small nonprofit that fine-tunes open models on books it borrowed enters the same zone of risk as a multinational, only without the legal budget. This is what Wikimedia means when it fears copyright holders could gatekeep even innovative and transformative uses of their own work.
Wikipedia’s own tooling is in this blast radius. Editors currently lean on language models trained on broad corpora to flag vandalism, detect copied text and hunt for gaps. A maximalist ruling invites challenge after challenge against the vendors of those tools, and the Wikipedia AI copyright brief lists these defenses as uses the foundation cannot afford to lose.
Open Source in the Crossfire
The collateral damage spreads through the software world. LaPorte’s statement warns that a careless ruling could stifle the development of open source technologies, and open models trained on mixed corpora are precisely where licensing diligence is weakest and experimentation is strongest. If only giant incumbents can clear the fee barrier, the ruling would hand them a regulatory moat dressed up as moral clarity.
| Outcome | Big AI labs | Small builders and nonprofits | Free knowledge sites |
|---|---|---|---|
| Publisher victory | Pay once, keep the moat | Priced out of training data | Lose AI editing tools; traffic still gone |
| AI victory | Status quo forever | Fine, at first | No attribution, no licensing leverage, funding starves |
| Wikimedia middle ground | Train free, pay at the output boundary | Training stays legal, outputs stay simple | Tools survive, credit and license flow returns |
Publishers, for their part, would argue the gatekeeping fear is theoretical while the copying was documented. That argument deserves the hearing it gets in court, not the silence of a settlement drafted around small builders. That is the other half of the Wikipedia AI Copyright bargain.
What an AI Victory Would Break
When Licensing Becomes Optional
Now run the Wikipedia AI Copyright board the other way. A ruling that declares both scraping and output fair use, at any scale, is the green light Wikimedia warns of: permission to ignore licensing terms entirely, because terms become unenforceable once the use they govern is free by judicial fiat. Licensing markets do not recover quietly from that. They dissolve into a land grab in which only whoever holds the largest corpus and the largest distribution counts.
The fragmentation warning hides inside the word flow. The open web survives because authors, publishers and audiences keep exchanging something: credit, traffic, payment, permission. A ruling that lets machines take the knowledge and skip every exchange fragments the flow of knowledge between human authors and audiences, which is exactly the phrase the Wikipedia AI copyright brief chose.
Communities Starve Without the Loop
For Wikipedia the arithmetic is personal. Articles exist because readers become editors, and readers arrive through search and citation. When answers arrive fully formed, with neither credit nor a reason to visit, the pipeline that converts curiosity into volunteer labor dries up. The Wikipedia AI copyright position is that substitution without attribution is not a neutral technological fact but a policy choice about whether communities like this one get to continue.
Wikimedia’s CEO Bernadette Meehan framed the alternative as a shared duty: expanding access to knowledge through AI is a powerful goal, but it must respect the people, contributors, and communities that make reliable knowledge possible.
There is also a slower, subtler failure mode in the publisher-victory scenario: self-censorship. When the cost of a lawsuit is measured in millions and the cost of licence compliance in thousands, risk-averse organisations stop experimenting altogether. Much of the early energy around small, locally focused AI assistants came from open datasets, and a Wikipedia AI Copyright loss would chill exactly that kind of low-budget experimentation for years.
Why the Wikipedia AI Copyright Brief Matters to You
If You Build With AI
Whatever Wikipedia AI Copyright verdict lands will set your compliance weather. Build models or apps and the middle ground is the most workable forecast: train broadly without permission requests, then keep outputs far enough from substitution that licensing questions stay rare and cheap. It also affects the tools you trust. As sites install more barriers, our report on AI agents being blocked by websites shows the ecosystem already bracing for a less permissive web. Courts may decide whether that brace becomes permanent law.
If You Publish Anything
If you publish, the fight decides whether your work enters machines with your name attached or without. The scraping backlash our coverage of Meta delaying its AI scraping terms captured is one half of the story: platforms negotiating permission before the law does. The Wikipedia AI copyright brief is the other half: communities asking the courts to make permission matter after the fact, at the moment a machine speaks your words back.
If You Just Read
Everyone who reads is a participant in whichever ecosystem the ruling selects. A web where free knowledge gets credit and traffic funds the next edit produces sources worth citing. A web where substitution wins produces confident answers with no authors behind them, and no one left to correct them when they are wrong.
The Wikipedia AI copyright filing will not decide these lawsuits by itself. Its value is that it puts a working middle on the table at a moment when both edges look like cliffs.
Finally, remember that these outcomes are not written in stone. Court rulings set boundaries, but licences, fundraising and public expectation fill in the rest. The same community that rallied around Wikipedia through paywall experiments and SOPA protests turned out for the Wikipedia AI Copyright debate within days of the amicus filing, which tells you that users understand the stakes even when the legal language is dense.
References
Wikipedia’s operator says a bad AI copyright ruling could threaten the open web (The Verge)
The New York Times v. OpenAI and Microsoft complaint (December 2023)
OpenAI and journalism (OpenAI statement)
Microsoft’s September filing in the Times authors lawsuit (The Verge)
Even Wikipedia is hemorrhaging traffic to AI (The Verge)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
