The material a model learns from
Training data is the large body of material used to train a model — text for an LLM, images for an image recognition system. Models become capable by finding patterns in it, so what was learned, and how much, determines both character and capability.
Quality matters as much as volume
More data helps, but quality matters too. Train on inaccurate text and the model learns the inaccuracies; train on a skewed sample and the skew shows up in its judgments as bias.
Face recognition trained mostly on one demographic performing poorly on others is the classic lesson that a bias in the data becomes a bias in the model.
What LLMs are trained on
Models like ChatGPT learn from web text, books, encyclopedias, and program code — an enormous corpus that inevitably includes outdated and incorrect material, which is one reason answers are not automatically right. There is also a cutoff date beyond which the model has no knowledge of events.
Copyright and privacy questions
Whether text and images on the web may be used for training without permission remains contested worldwide, alongside questions about personal data. See the entry on AI and copyright for more. Knowing this background makes both AI news and changes to service terms easier to interpret.
As human-created data grows scarce, synthetic data generated by AI is increasingly used for training. Securing high-quality data has become one of the decisive factors in the competition between labs.
When an answer seems dated or slanted, recalling the scope and skew of the training data helps you read the limitation calmly.