What this is
The model is a file of numbers, called weights, left from training. It knows what it read while it was being trained, plus whatever is pasted into the conversation. It does not know an inbox, a company's records, or what a government database says today.
Text pasted into the thread sits in a context window: a limited working space for this conversation. Past the edge of that window, the earlier turns are gone.
Those are the only two sources unless another system fetches a document or calls a tool. A fluent answer about a contract or a docket can still be a guess. Fluency is not evidence that the model saw the file.
A model is a frozen set of parameters — weights — produced by training. At inference those weights do not update against the world.
The context window is ephemeral working memory for the current session. Tokens that fall outside it are not available.
Parametric knowledge plus in-context tokens are the only sources unless some other system retrieves text or calls a tool. Fluency is not evidence of access.
Easy to mix up
This is easy to mix up with a system that can see private records. Access is a separate job. If nobody fetched the file and put it in the window, the model did not see it.
It is also easy to mix up with memory. The weights are a snapshot from the end of training. The window is this conversation only. Neither updates itself against an inbox, a docket, or a database.
Live access to private systems is not a property of the generator. It is a property of retrieval and tools attached later.
Staleness is the natural state of a file that stopped learning. Session context is not long-term memory.
From work
A clause from a vendor contract is pasted into chat with a question about last year's master agreement. The model can write a confident comparison. It has the clause. It does not have the master agreement unless that went in too. The comparison is a plausible MSA-shaped paragraph, not a reading of the file on the shared drive.
If that comparison is pasted into an email to counsel, the guess has left the chat. The model completed a sentence. It did not read the drive.
A pasted clause is in-context. An unpasted master agreement is not. A generated comparison is a sample from the prior, conditioned on the clause and on whatever MSA-shaped text lived in the weights.
Downstream use — an email, a markup, a filing — is a human act. Hallucination here means an ungrounded span, not a moral failing of the model.
What to ask next time
Was this in the weights, in this window, or fetched from a system we trust? If it was not fetched, treat names, dates, and “the contract says” as invented until a document is in view.
Ask which spans are parametric, which are in-context, and which were retrieved. Ungrounded proper nouns and dates should be treated as samples until a source is attached.