Prologue

You can learn a lot about an institution from the rules it creates and the way it chooses to enforce them, or not.

In March 2024, the Wikimedia Foundation admitted in writing that many of the AI companies using Wikipedia might be violating its license. That license carried the promise the whole project rested on: what is built from shared work stays shared. The Foundation enforced nothing, and its commercial arm kept collecting from the industry it suspected. In June 2026, the same institution expelled the man who had co-founded Wikipedia and given it its name. His offense, in the end, was disagreement.

To understand those two choices, it helps to begin with a copy.

On October 27, 2025, Elon Musk’s xAI launched a website called Grokipedia. It was a near-complete copy of Wikipedia, produced by running the encyclopedia through Musk’s AI and rewriting it to correct what Musk and many others saw as systematic bias in Wikipedia’s sources and framing. At launch, though, the AI “balanced” that bias by pulling in sources Wikipedia’s own rules excluded, among them InfoWars and neo-Nazi forums.

The motivation deserved a hearing. The worry that Wikipedia had developed a political lean, in its articles and in its rulings on which sources counted as reliable, was shared by many readers and critics. The execution deserved none. An AI set loose on a reference work, correcting it against the open internet, produced exactly the contamination Wikipedia’s standards existed to prevent. It was crude and easy to condemn, and the Foundation condemned it.

But condemnation was all the Foundation offered. The problem was the Foundation’s passivity, not Musk himself. Musk had merely done in public what others had been doing in private for years. For four years the Foundation had signed deals with AI companies that trained on Wikipedia without meaningful attribution. It had never sued OpenAI, whose chatbot leaned on Wikipedia more than any other source. It had never demanded that Meta release the trained model itself under the same license that governed the text Meta trained on. The Foundation’s own staff had admitted, in a March 2024 email, that many of these companies “may not be compliant with the letter of the Creative Commons rules or the spirit of the licenses.”

It kept selling to them anyway.

There was an uncomfortable irony buried in the outrage. In purely structural terms, Musk’s copy was more compliant with Wikipedia’s license than the companies the Foundation served. It credited Wikipedia and offered free public access. OpenAI did neither, and paid nothing. Grokipedia’s neo-Nazi sourcing was a real and serious failure of content, but a failure of the visible kind that scrutiny can force out, and within months it mostly had: later versions scrubbed the most notorious citations. The deeper harm was the one no one was protesting: the quiet absorption of Wikipedia into proprietary systems that could not be examined, corrected, or forked at all.

That harm had a history. In 1982, a company called Symbolics took code Richard Stallman had helped write, improved it, and refused to share the improvements back. He answered by inventing copyleft: a license built to keep free work free. Anyone who improved the work had to share their changes under the same terms. Wikipedia launched in 2001 under a license of that kind. The promise was simple: use this freely, but keep it free.

Twenty-four years later the promise had eroded, and not mainly at Musk’s hands. OpenAI, Anthropic, Google, and Meta absorbed Wikipedia into proprietary models, gave no attribution to end users, charged for access, and licensed their outputs on their own terms. Community work absorbed, proprietary value added, others locked out: Symbolics again, at industrial scale.

The institution built to prevent this collected more than $180 million a year. It employed some 700 staff and contractors. It ran a commercial arm, Wikimedia Enterprise, that sold structured data access to Google, Amazon, and Microsoft.

And when the moment came to defend the commons its volunteers had built, it issued a press statement. The statement noted that “no single individual, company, or agenda can exert influence over the work.” It did not threaten legal action. It did not assert a licensing violation. After four years of accommodating the invisible extraction, it had forfeited the credibility to single out the visible fork.

This is not a story about Elon Musk, or even about artificial intelligence broadly. It is a story about one of our most important institutions. A volunteer movement to free the world’s knowledge became a nine-figure foundation that sold access to that knowledge’s extractors, then would not defend the commons when it mattered. The promise was broken, not by an outside force, but by the institution meant to uphold it. The antagonist is not a person. It is a failure: of nerve, of independence, of the will to enforce a promise.