Training Data
Training data is the collection of examples a model learns from. For a language model it might be billions of pages of text; for an image model, millions of pictures. The model studies this material to pick up patterns, so the quality, variety, and fairness of the data shape everything the model can do. As the old computing saying goes, garbage in, garbage out: a model is only ever as good as what it was taught with.
A model does not arrive knowing anything. Everything it can do, it learned by working through examples, and that body of examples is its training data. Show a system millions of photographs labelled with what they contain, and it learns to recognise those things. Feed a language model a huge sweep of books, articles, and web pages, and it learns the patterns of how people write. The data is not a minor ingredient; it is the raw material from which the model’s abilities are built.
Because of this, the choice of training data has enormous consequences. If the data is narrow, the model will be too. If it is riddled with errors, the model will absorb them. Most importantly, if the data reflects human biases, whether about gender, race, or anything else, the model tends to reproduce and sometimes amplify those biases. A hiring tool trained mostly on past decisions from a biased industry will learn to be biased in the same way, which is why data quality is a genuine ethical and legal issue, not merely a technical one.
Training data is also at the heart of some of the biggest debates in AI. Where did the data come from? Did the people who created it agree to its use? These questions drive live disputes over copyright, privacy, and consent, and they feature heavily in the UK’s evolving approach to AI regulation. When you see an AI story about lawsuits, bias, or trust, training data is very often what the argument is really about.