You upload a document and receive an AI-generated summary. Did the system just train on your document? That question cannot be answered simply because the model used the document to produce an output.
Google’s machine-learning glossary distinguishes training, which determines learned model parameters, from inference, which uses a trained model to produce predictions or outputs. In a language-model application, generating a response is an inference task.
Separate the current response from later data use
A service may place your document in the current input, retrieve passages from a stored collection or use another workflow. That does not by itself tell you whether the provider retains the material or later uses it in training. Those are questions for the service’s actual settings, contract and data policy.
For a fictional document assistant, imagine that a staff member supplies a paragraph for summarisation. The model can use those words in the response without the organisation running a training job for that paragraph. If the organisation later fine-tunes a model on a curated dataset, that is a separate process to document and evaluate.
Ask which part a cost estimate describes
A training estimate may concern preparing data and updating a model. An inference estimate may concern serving requests at a particular volume and response length. Comparing the two as if they were prices for the same operation can mislead a project plan.
Also distinguish a changed prompt or retrieved document from a changed model. An application’s answers can change because its context or tools changed, even when the model version stayed the same.
- Which model version is serving the request?
- What information enters the current context?
- Is any separate training or fine-tuning process taking place?
- What retention and data-use terms apply?
These questions create a clearer technical conversation than asking whether the system “learned” something in an everyday sense. The word can refer to several different behaviours. Name the operation you mean, then test and document that operation rather than assuming a fluent answer reveals how the system was built.