Data bias

Data bias is when the training data over-represents some groups, perspectives, or patterns — and the model learns that skew, then reproduces it.

Because models compress their data, whatever’s over- or under-represented in the source gets baked in.

• If most historical text is from one culture, the model leans that way.
• If certain professions are stereotyped in the source, the model may echo it.