The training data is the world

A language model learns from text — and text reflects the world as it has been written about, not the world as it is.

Written text overrepresents certain languages, certain perspectives, certain demographics. It overrepresents people who write publicly, who have internet access, whose ideas circulate widely. It underrepresents oral traditions, non-Western intellectual frameworks, languages with less digitized text.

The model absorbs these imbalances. Not as explicit beliefs, but as statistical patterns. Certain names are more likely to be associated with certain professions. Certain dialects are treated as more "correct" than others. Certain cultural references feel natural; others require more effort to generate.

This isn't a problem that can be fully fixed after training. You can add fine-tuning data to correct specific biases, but the distribution of training data is foundational. A model trained on a biased corpus has learned a biased version of the world. What you can do is use it with that awareness, and push back when its defaults don't serve you.