NotebookLM and source-grounded assistants

Everything so far answers from the whole internet’s patterns. This lesson’s tool does the opposite — and the opposite turns out to be exactly what students, researchers, and professionals often need: an assistant that knows only what you handed it.

The fact sheet (community-verified)

Grounded means bounded

Give NotebookLM your sources — reports, papers, notes, transcripts — and ask questions. Answers come only from those sources, with citations that click through to the exact passages. Ask something the sources don’t cover and the well-behaved response is “your sources don’t say” — the sentence E1·L7 taught you that general assistants struggle to produce.

Why this matters: the hallucination problem (plausibility without a truth check) shrinks dramatically when the model is confined to a small, known pile of text and every claim carries a pointer you can check in seconds. It doesn’t vanish — passages can still be misread or stitched imperfectly — but verification collapses from “search the world” to “click the citation”.

This idea has a name you’ll meet properly in the Practitioner track: retrieval — and NotebookLM is its friendliest on-ramp. The audio overview feature (two AI voices discussing your sources like a podcast) is a genuinely useful review tool; treat it as a tour of your documents, not a replacement for them.

Where grounded tools win

Study packs (your lecture notes + the textbook chapter), synthesising a pile of reports, onboarding into a project’s documents, “what does the contract actually say” questions. Where they don’t: anything requiring knowledge beyond your sources, general drafting, and creative work — the general assistants’ home turf. One privacy habit before uploading anything sensitive: check the tool’s data-handling terms (the card links them), and your workplace’s policy (E2·L6 echo).

Try it now (5 minutes)

Upload two related documents you actually have (or any two public PDFs on one topic). Ask: “Where do these two sources disagree, and what does each say? Cite passages.” Click one citation. That loop — question, cited answer, one-click check — is grounded AI at its best.

Check your understanding

1. You ask a grounded notebook something its sources don't cover. The well-designed behaviour is: Answer from general knowledge without mentioning it Say the sources don't cover it Invent a plausible citation

2. Grounding changes verification because: No checking is ever needed again Citations point into your own documents, so checking a claim takes seconds The model becomes incapable of misreading

Recap

Source-grounded tools answer only from what you provide, cite into it, and say so when sources are silent — retrieval’s friendliest introduction. Next: the lesson that saves you money — what free and paid tiers actually buy, and when upgrading is worth it.