Whose voices shaped the AI you're using?
By Rodney Noriega, Founder, Universal Enablement · Published Oct 12, 2026
Short answer
AI assistants learn from huge amounts of online text: websites, books, Wikipedia and forums like Reddit. That material over-represents English speakers, North Americans and young men. Since AI learns values along with language, its answers lean the way its data leans. This applies across ChatGPT, Claude, Gemini and others, though few companies publish full details.
The clearest public example is the model behind the original ChatGPT. Its creators published the recipe in 2020. About 60% came from a large crawl of public websites. A big share, 22%, came from web pages linked on Reddit posts with at least three upvotes, a filter meant to signal quality. Wikipedia made up 3%.
That Reddit filter put young, male, mostly American users in charge of deciding what counted as "quality." Wikipedia's editors are also overwhelmingly male. And English makes up far more of online text than its share of the world's speakers.
Newer models use bigger and more varied data, but the companies share less about it. The basic pattern hasn't gone away.
For a Metro Vancouver business with customers from many cultures and languages, this shows up in real ways. AI drafts may assume North American norms, miss cultural context, or work less well in some languages. Test any workflow on real examples from your actual customers before you rely on it.
Related
Start with one workflow. In 30 minutes we'll look at where your team's time goes and tell you honestly whether AI can help.
Book a free 30-minute workflow review