What a chatbot knows, sees, and looks up

Four sources that are easy to confuse

Imagine asking for tomorrow's school timetable. A model's training is not a reliable source for your personal timetable. The application needs relevant information. We will follow four ways information can be involved.

Learned parameters: training shapes the model's behaviour. This can support broad knowledge and language skills, but is not a searchable notebook of guaranteed facts with citations attached.

Current context: the messages and other material supplied for this response. If you paste a timetable, the system can answer using it. The information came from your prompt, even if it was never in training.

Saved product memory: some applications retain selected information across conversations. Whether this exists and how it works depends on the product and settings. It is not the same as retraining the underlying model.

Tools and retrieval: an application may search, read a file, query a database, or run code, then provide the result to the model. That capability must actually be available and used. A confident sentence saying “I checked” is not, by itself, evidence of a successful check.

A worked source-based answer

You provide this invented notice:

Science club meets Thursday at 3:30 in Room 12. Bring a notebook. The notice does not give an end time.

Then ask:

Using only this notice, when and where does science club meet, and when does it finish? If the notice does not say, say that.

A supported answer names Thursday, 3:30, and Room 12, then says the end time is not supplied. An answer of “It finishes at 4:30” adds information not in the source. The right correction is to remove or verify the extra claim, not to improve its wording.

Why a follow-up can work

In the same conversation, “Put that in a checklist” can refer to the notice and answer already in context. In a new conversation, “that” may have no referent. Supply the material again when needed.

Conversations also have finite context. An application may shorten or summarise old messages or retrieve only part of a file. When an answer misses a condition, restate the relevant passage and ask the system to identify the evidence it used. Do not assume every uploaded page was read equally thoroughly.

Use this distinction

For a current opening time, ask for a current source and open it. For an arithmetic result, use a calculation tool and inspect the inputs. For a summary of your document, supply the document and check that the summary preserves its claims and uncertainty.

The goal is to know where an answer could have come from. “The AI knows it” is too vague to tell you whether the answer is supported.

Trace the source of each sentence

Suppose a school-club assistant says: “Our meeting is Thursday. Bring two pencils. The library closes at six.” You told it the meeting day in the current chat, pasted a fictional packing note mentioning pencils, and supplied nothing about library hours. The first two claims have identifiable conversation sources. The third may come from an unsupported continuation unless the tool actually retrieved an appropriate source.

Ask what information was available for each claim: the current prompt, earlier messages still in context, a file, a retrieved page, or patterns learned during training. Do not treat a citation-shaped string as proof of retrieval. Inspect whether a source was actually supplied and whether it supports the particular claim.

Now change the fictional day to Friday in a follow-up. A useful assistant should apply your correction within this conversation. That is a context update, not evidence that its original training weights have been rewritten.