What is training data?
Training data is the data an AI model has learned from, such as text, images or sound.
What the model can do, and what it misses, comes from its training data. A language model trained mostly on English text is often better at English than at Swedish.
What has been used for training is hard to remove afterwards. Information that has become part of a model cannot be deleted like a file.
That is why it matters whether your documents, questions and answers are used to train a supplier's models. The answer should be in the contract.
How it works
A large language model is trained in stages. First on large amounts of general text, often from the web, books and code. Then on smaller, carefully chosen examples that teach it to follow instructions and answer helpfully. Both stages shape what the model can do and how it expresses itself.
Why it matters
The training data decides which languages, subjects and perspectives the model handles well. It also decides which skews come along. That is why it is worth asking what a model was trained on, and testing it on your own examples before it is used in important workflows.
What to keep in mind
Training data has a cut-off date. The model knows nothing about what happened or changed after it, unless it reads current documents when it answers.
Under the EU AI Act, providers of general-purpose AI models must publish a summary of the content used to train them.
What it means for you
Require a clear statement in the contract that your data is not used to train models.
Common questions
Is AI trained on our documents when we use it?
With RAG, the model reads the documents when the question is asked and no training takes place. Some services do use customer data for further training, so check the terms.
Why does training data matter?
A model reflects what it learned from. Skews in the data can lead to bias, and the model knows nothing of what happened after a certain point in time.
Can training data contain personal data?
Yes. Models trained on large amounts of text may have learned information about people. GDPR then applies to the training too.
Related terms
Read more
All terms Updated