What it means
Everything a model knows comes from its training data. For large language models that means an enormous mix of web text, books, code, and licensed material, increasingly supplemented with synthetic data generated by other models.
Composition matters more than volume. A model trained mostly on English performs worse in other languages; one trained heavily on code reasons differently; one trained on internet argument absorbs internet argument. Because the patterns are learned rather than specified, whatever is in the data comes along — including the parts nobody intended.
Training data is also the field's central legal fight. Whether training on copyrighted material without a license is permissible is being litigated in several jurisdictions, and the answer will shape which models can be built and sold.
Why it matters
Data provenance is a business risk, not just an ethics topic. A model trained on material a vendor had no right to use is exposed to injunctions and damages, and so, potentially, are its customers. It's also why vendors increasingly advertise licensed or indemnified training data.
In practice
For high-stakes deployments, ask what a model was trained on and whether the vendor indemnifies you against training-data claims. For quality, expect performance to track how well-represented your domain and language were.