Training data is the data an AI model learns its patterns from. For a language model, it is vast amounts of text from the internet, books and other sources.
The model does not store the texts as an archive. It learns statistical relationships between words and concepts and uses them when it writes. That is why its answers reflect the training data: its language, its knowledge and its skews. If Danish is sparsely represented, the model's Danish is correspondingly less precise, and skews in the data can reappear as bias.
Training data also has a cut-off date. The model only knows about events after it if it is given access to new sources, for example through search or RAG.
In practice
When a colleague asks whether the AI knows anything about your organisation, the answer is usually no. It knows patterns from its training data. So give it your own sources as part of the task rather than relying on it to know you.
Training data also raises questions of copyright.
The concepts in this glossary run through my AI courses and workshops, where we work with them on your own tasks.