Understanding an AI's knowledge
Large language models don't hold knowledge or awareness the way people do. They generate responses from patterns learned during training, and the data they were trained on is the basis of what they "know".
What AI is trained on
Models are trained on large and varied collections of text — a broad cross-section of the internet, books, and other written material. Training datasets can include public information about people; they are not necessarily free of personal information.
A pretrained model does not automatically have access to your school's private records. Sparky can use school information and materials made available to it in Flint during a conversation. That supplied context is separate from the data used to train the underlying model. For Flint's handling of student data, see Communicating with parents about AI.
Knowledge cutoffs
Every model has a knowledge cutoff: the point in time where its training data stops. If a model was trained on data up to a certain month, it has no built-in awareness of anything that happened in the world after that.
This is why a model can be confidently wrong about recent events, and why it may not know about something released after it was trained.
What Sparky can look up
A knowledge cutoff isn't the whole story, because a model can be given information it was never trained on. Sparky can:
- Search the web during a conversation, and cite the sources it used, so it can answer questions about recent events.
- Check the current date and time, so it isn't guessing about "today".
- Read the materials you attach — a worksheet, a textbook chapter, lecture slides, a link — and answer from those rather than from training data. See Uploading content to Flint.
- Look up AI policy documents your school has uploaded, so you can ask what your school's policy says. This is document retrieval during the conversation, not model training.
For the full list, see Tools available when chatting with Sparky.
What Sparky remembers
A language model on its own starts every conversation from nothing. Flint adds a layer on top of that:
- Within a conversation, Sparky uses recent messages and, in a long conversation, Flint's summary of older messages. A summary can omit details, so repeat exact wording if Sparky loses it. Summarizing does not delete the messages from your chat history. See What is a context window?.
- Between conversations, Sparky can save short Memories — things like a student's interests, preferred explanation style, or grade level — and use them to personalize later responses. Memories are saved per school and do not carry from one school to another. See Memories in Flint.
Memories are deliberately small, and they belong to the person they describe: you can open yours, read every one, and delete any you'd rather Sparky didn't keep. For how Flint handles student data more broadly, see Communicating with parents about AI.
Key points to remember
- A model's built-in knowledge comes from its training data and stops at its knowledge cutoff.
- Training data can include public information about people. A model does not automatically have access to your school's private records.
- Sparky can search the web, check the time, and read the materials you attach, so it isn't limited to what it was trained on.
- Sparky keeps a small set of Memories between conversations, which you can review and remove.