The Hidden Database Behind Wikipedia Is Skewing What AI Knows

Most people know Wikipedia, the online encyclopedia anyone can edit. Far fewer know about Wikidata — its quieter sibling. Instead of articles, Wikidata stores plain facts about people, places, and things in a format computers can read directly. AI companies, search engines, and chatbots increasingly treat it as a source of truth when answering questions about the real world. That makes what is (and isn't) in it surprisingly powerful.

A new audit by researcher Clair Kronk checked roughly 10 million facts about more than 6 million people in Wikidata. The results show large holes. Women make up just 28.71% of the people with a listed gender. Only about 38% of people have a citizenship recorded at all, and of those, more than half come from Western Europe and North America. Among the 1.8 million birthplaces that could be classified, rural areas accounted for 2.48% — compared with about 27.4% of the actual global population. Ethnicity is listed for under 1.2% of people, and fewer than one in five entries has a description in any language other than English.

Why should you care? Because when you ask an AI assistant "who are the most important scientists?" or "what does a typical family look like?", the answer is shaped by whatever data sits underneath. If the underlying list over-represents wealthy, urban, Western, English-speaking people, the AI's picture of the world tilts the same way — quietly, and without telling you. That can affect everything from search results to hiring tools to how companies decide what customers want.

There is an honest catch here. Missing information is not the same as wrong information. Wikidata not listing someone's ethnicity may reflect privacy choices, gaps in sourcing, or simply that no volunteer has gotten around to it. The study is a first measurement, not a fix — and it's published as an open-source tool so others can re-run it. Filling these gaps takes volunteer time, and there's no guarantee AI companies will use a corrected version. Still, you can't repair what you haven't measured.