Page 02 / Context
In 2005, Nature asked whether Wikipedia could match Encyclopædia Britannica. The result became one of the defining media moments of the internet age: the open, volunteer-written encyclopedia was found to come unexpectedly close to the accuracy of the most prestigious reference work in the English-speaking world. Wikipedia was no longer just the chaotic child of the web. It had passed the gatekeeper's test.
For the next two decades, Wikipedia became one of the default interfaces of public truth. It was there in the extra browser tab while people wrote emails, essays, articles, school reports, policy notes and arguments. It made verification feel instant. It compressed what once required shelves, indexes, libraries and institutional access into a reflex: open tab, search, check, return.
Wikipedia publishes in more than two hundred languages. Each edition is maintained by its own community of volunteers, working independently. What happens in one language edition is not automatically visible in another. The communities do not all read each other. They cannot — no one speaks all those languages.
That also means no one has ever been able to read Wikipedia as a whole. Not its editors, not its founders, not its critics. The complete archive has always been too large, and too multilingual, for any individual or institution to hold in view at once.
In 2005, Nature asked whether Wikipedia could match Britannica.
We ask a different question: can Wikipedia match itself?
We returned to the forty-two subjects used in Nature's original comparison — and instead of measuring Wikipedia against an outside source, we compared its language editions with one another.
That kind of comparison has only recently become possible. Articles can now be collected, translated, aligned and checked across languages at scale. The tool that makes this possible is also the one Wikipedia, in 2026, voted to keep out of its editorial process.
Look at what happens when the same subject is opened across languages.
Page 03 / Three exhibits
These panels are designed as entry points, not verdicts. Each subject should lead to the original language edition, the extracted sentence, and the aligned comparison layer. The chart is the door. The source sentence is the evidence.
Replace chart points with final extracted source sentences before publication.
Use with care: not every framing difference is a contradiction.
Final exhibit should be selected from the cleanest verified evidence.
Page 04 / Findings
What emerges is not simply a list of mistakes. Across languages, Wikipedia looks less like one encyclopedia translated into many versions than like a set of related editions, built by different communities, in different languages, with different habits of emphasis, omission and correction.
Some differences are minor. Some are cultural. Some may be translation. Some may be editorial inheritance. Some may be ordinary error. But some touch the kinds of facts encyclopedias are supposed to stabilise: dates, quantities, definitions, technical units, classifications, causes and chronologies.
Wikipedia remains one of the most useful public knowledge systems ever built. But this audit suggests it should not only be cited as a source. It should also be studied as an object.
The historical perspective is difficult to avoid. Wikipedia once became powerful because it made knowledge searchable, public and fast. Now the next knowledge interface makes something else possible: not just asking Wikipedia questions, but reading Wikipedia itself across languages.
Britannica, the institution Wikipedia once displaced, has already begun to experiment with that shift. Its AI answers come with a revealing warning: "AI-generated answers from Britannica articles. AI makes mistakes, so verify using Britannica articles." The old encyclopedia is trying, however awkwardly, to become an encyclopedia that answers back.
Wikipedia has taken the more defensive path. At the moment when the tools exist to inspect its full multilingual shape, it has voted to keep those tools out of its editorial process. That caution is understandable. But it also means the technology capable of seeing Wikipedia as a whole is treated first as a threat to Wikipedia's parts.
The claim is narrow. It is not that Wikipedia is false. It is not that every difference is an error.
That changes the hallucination debate. Large language models do not enter a clean archive and corrupt it from the outside. They also inherit the inconsistencies, gaps and local histories of the public knowledge systems on which they are trained.
The machine did not invent the fracture. It made it easier to see.