NotebookLM launches as an experimental AI notebook
By restricting the model to only the documents a user uploaded, Google aimed to cut hallucinations — an early mainstream deployment of what became known as retrieval-grounded AI.
- Models & capabilities
- Minor
Google Labs opened waitlisted early access in the United States to NotebookLM, an experimental AI-powered research and note-taking tool that had been previewed at Google I/O two months earlier under the name Project Tailwind. Users could upload their own documents — notes, PDFs, articles — and the tool would summarise them, answer questions grounded in the uploaded material, and help brainstorm or draft new text based on it.
The defining design choice was source-grounding: rather than drawing on the model’s general training data, NotebookLM restricted its answers to only the material a user had chosen to upload, with citations back to the source. Google presented this as a way to reduce hallucination — the tendency of language models to generate fluent but incorrect or invented content — while also giving users a degree of control over what the model could see, addressing privacy concerns that came with feeding personal documents into a general-purpose chatbot.
NotebookLM remained a Google Labs experiment through its first year, but the underlying idea — constraining a language model to a bounded, user-supplied document set rather than its full training corpus — became one of the more durable interface patterns to come out of the post-ChatGPT wave, later adopted widely under the label retrieval-augmented generation. The product continued to be developed and renamed in subsequent years, becoming Gemini Notebook in mid-2026.