Whose world?

Whose internet?

In chapter 7, you saw that the pictures and answers of a model come from its training data. A lot of that data comes from the internet. So which languages fill the internet?

The website W3Techs keeps track of the languages on the 10 million most visited websites. In August 2026, 49.5 percent of them were in English. Spanish and German came next, with about 6 percent each. Swedish had 0.5 percent.[1]

Many AI companies use Common Crawl, a non-profit organization that copies billions of web pages and gives them away for free. Of 47 language models from 2019 to 2023, at least 64 percent were trained on its data.[2] In one Common Crawl collection from December 2023, about 44 percent of the pages were mainly in English.[3]

The training data can be even more English than the web. When the company Meta trained its model Llama 2 in 2023, almost 90 percent of the data was in English. Only 0.15 percent was in Swedish.[4] But look closer. In the same data, there was more Swedish than Chinese or Spanish, which had 0.13 percent each. Swedish is a small language, but it does quite well online. Much bigger languages can have less.

Some people are not online at all. According to ITU, the UN agency for digital technology, about 6 billion people used the internet in 2025, and about 2.2 billion did not. Almost all of the people offline, 96 percent, live in low- and middle-income countries. In rich countries, 94 percent of the people use the internet. In the poorest countries, only 23 percent do.[5] This gap is called the digital divide. When people are not online, their words, songs and stories are not in the training data either.

Who wrote it down?

You met the researcher Joy Buolamwini in chapter 7. In 2018, she made “AI, Ain’t I A Woman?”, a spoken word poem on video. It is less than four minutes long. She shows photos of famous Black women, such as Oprah Winfrey, Serena Williams and Michelle Obama. Then she shows what face analysis systems from big tech companies said about them. Often, the systems labeled the women as men.[6] Watch and listen before you read on. What do you see and hear?

The title comes from a speech. Sojourner Truth was born into slavery in the state of New York, and her first language was Dutch.[7] She learned English, but she never learned to read or write.[8] In 1851, she gave a speech at a meeting for women’s rights in Akron, Ohio. Everything we have from the speech was written down by other people, and there are two famous versions.[9]

Most people today know the second version, and Buolamwini’s poem borrows its famous question. But Truth did not grow up in the South, and Gage’s version also says that she had 13 children. She had five. Historians such as Nell Irvin Painter think that the first version is closer to what Truth really said.[7]

Why does this matter for AI? A model learns from what people wrote down and copied. When other people write about you, their version is the one that ends up in the training data.

Whose values?

In chapter 7, you could ask a chatbot about driving licenses in English and in Swedish. You also read about a study where three out of four answers in English were based on rules in the USA.[10] Rules are one thing. What about values, like what is important in life?

In the World Values Survey, people in many countries answer the same questions about their values. In 2023, researchers asked a GPT model from OpenAI, the company behind ChatGPT, the same questions. The answers were most like the answers from people in the USA and other WEIRD countries. The further a country’s culture was from the USA, the less the answers looked like the answers from that country.[11]

In 2024, another group tested five models from OpenAI. All of them gave answers with values like those in English-speaking countries and in Protestant countries in Europe. Then the researchers told the models to answer like a person from a certain country. With the newest models, this made the answers closer to the real answers in seven or eight out of ten countries.[12]

Sweden is one of the Protestant countries in Europe. So for many people in Sweden, the values in the answers may feel normal. That does not make them neutral. A bias is easiest to miss when it looks like your own world.

Now test it yourself with everyday life. Use the chatbot that your school has chosen. Ask a question in English, and then the same question in Swedish in a new chat. Ask each question two or three times, because the answers change. Which country and culture does each answer start from? Do you know a third language, from home or from school? Then add it. You are the expert, because you can judge if the answer sounds right.

How do people celebrate when they finish school?

In Swedish: “Hur firar man när man slutar skolan?” You can also choose your own everyday question, for example about breakfast, birthday presents or a normal weekend.

AI for more languages

Many languages have much less data than Swedish. They are called low-resource languages, and people around the world are working to change that.

Africa. Masakhane is a community for language technology in African languages. The name means “we build together” in isiZulu, and the idea is technology for Africans, by Africans.[13]

Sweden. In February 2026, the Wallenberg Foundation and researchers from eight Swedish universities started work on a large language model for Swedish. Authors, publishers and news media will give texts that editors have checked. The Prime Minister presented the project.[15] In May 2026, the government asked a national coordinator to help Sweden get language models that reflect Swedish culture, society and law. She will also look at how authors and others who own texts can be paid when their texts are used to build language models.[16]

Everyone. In December 2024, Common Crawl asked people who speak languages other than English to send in the addresses of websites in their languages. Then more of those websites can be copied.[17]

All of this costs money. Who should pay: companies, countries, foundations or the people who speak the language? People disagree.

Final task: whose world is in the answer?

Do your own investigation, and present it in English. Choose a question about everyday life, traditions or values, or about a famous person or event. Ask it in at least two languages, three times in each language, in new chats. Save all the answers, even the boring ones.

Your investigation must have:

1. Your question. What did you ask, and why?
2. Your method. Which chatbot did you use, on which date, in which languages, and how many times?
3. Examples. Quote at least two short parts of the answers, word for word.
4. An explanation. Why do you think the answers turned out like this? Use the term training data.
5. Your reflection. What surprised you? What can your small test not show?
6. Advice. Give one piece of advice to someone your age who uses AI.

Choose one of two forms. Your teacher tells you which.

• Spoken: a talk of 3–4 minutes, live or recorded. Show your examples on a slide or on paper.
• Written: a report or an article of 350–500 words, with a title.