Data
Training data is the collection of examples an AI model learns from; the patterns in that data determine what the model can do and the mistakes it will make.
A model only knows what its training data showed it. For an image classifier, that means labeled photos. For a language model, it means trillions of words from websites, books, articles, and code. During training, the model adjusts itself to fit this data, so the data is, in a real sense, the model's entire education.
Quality matters as much as quantity. If the data is full of errors, the model learns errors. If it lacks examples of some group or situation, the model will be unreliable there. Teams spend enormous effort collecting, cleaning, filtering, and de-duplicating data before training even starts.
Training data also raises hard questions. Much of it is scraped from the public internet, which involves copyright disputes and privacy concerns. And whatever biases exist in the text people wrote will be absorbed by the model unless carefully corrected.
An image model trained on millions of captioned photos learns that 'golden retriever' goes with a certain fluffy shape, because it saw that pairing thousands of times.