NotebookLM answers only from the sources it is given

Grounding as the stated difference from a chatbot

Google's announcement of NotebookLM, dated 12 July 2023 and framed as an experimental release from Google Labs, states the design intent directly: 'A key difference between NotebookLM and traditional AI chatbots is that NotebookLM lets you "ground" the language model in your notes and sources'. In practice this means the model answers from documents a person has chosen to upload, initially Google Docs, rather than from open-ended training knowledge alone. The announcement also says each response arrives 'with citations, showing you the most relevant original quotes from your sources', intended to let a reader check a claim against the material it came from, and states that 'the model only has access to the source material that you've chosen to upload'.

A feature added over a year later: Audio Overviews

On 11 September 2024, Google's Audio Overviews announcement extended the same grounded model into a new format: two AI-generated hosts discuss the uploaded sources aloud, described as able to 'summarize your material, make connections between topics, and banter back and forth', with the resulting audio downloadable for offline listening. The mechanism is still source-bounded; the hosts are not meant to introduce outside material, only to narrate and connect what was uploaded.

What grounding does and does not guarantee

Google's own Audio Overviews post is unusually direct about the limits of the feature it announces. It states the hosts 'sometimes introduce inaccuracies' while explaining sources, that at launch 'they only speak English', and, most importantly for how the feature should be read, that the resulting discussion is 'not a comprehensive or objective view of a topic, but simply a reflection of the sources that you've uploaded'. That last point generalises to NotebookLM as a whole: source-bounded answering narrows where an answer can come from, which reduces one kind of error, invented facts from nowhere, but it does not remove the more basic risk that the uploaded sources themselves are incomplete, biased or wrong, or that the model still misreads or misquotes a source it was given. A citation lets a person check a claim quickly; it does not do the checking for them.

• Do the uploaded sources actually cover the question being asked, or only part of it?
• Has a sampled answer been checked against its cited source rather than accepted on the strength of having a citation at all?
• For an audio summary specifically, would a factual error be easy to notice without reading the underlying source first?

Grounding is a real constraint on where an answer can come from, and Google documents it as a deliberate design choice rather than a side effect. It is not, by the company's own admission, a guarantee of accuracy, which makes the citation the more important feature of the two.