grounded

On this channel, grounded appears in the context of NotebookLM — Google's AI research assistant co-created by author Steven Johnson — and it names the tool's core design principle: everything the model says must be tied to the documents you provide. You open a notebook, upload sources (journals, work documents, research quotes), and the model becomes an expert in exactly that material [^0cc0dd07-dc88-5d83-8cf7-2f6e48821796@612]. Johnson traces the idea back to the team's starting point: “we really started with the idea like let's build a model that is fully grounded and then we'll really just stick to the facts” [^0cc0dd07-dc88-5d83-8cf7-2f6e48821796@118].Grounding shows up as restraint. When asked to speculate about Johnson's personality from his saved quotes, the model “hems and haws” and declines to go beyond the sources [^0cc0dd07-dc88-5d83-8cf7-2f6e48821796@1377]; ask about something not in the notebook and it refuses to answer [^0cc0dd07-dc88-5d83-8cf7-2f6e48821796@1458]. Johnson says research mode deliberately errs toward groundedness over open-endedness [^0cc0dd07-dc88-5d83-8cf7-2f6e48821796@1458]. In practice, grounding also means inline citations: every answer shows exactly which source passage generated it, and clicking jumps straight to the original quote [^0cc0dd07-dc88-5d83-8cf7-2f6e48821796@1032].One boundary worth knowing: grounding applies to what the model answers from, not how the content gets there. Loaded copyrighted material sits in the model's context window — it is never used for training [^0cc0dd07-dc88-5d83-8cf7-2f6e48821796@841].