NOTEBOOKLM FOR HEALTH RESEARCHERS: WHAT IT IS, WHAT IT'S REALLY FOR, AND WHAT HAPPENS TO YOUR DATA.

By María García-Puente · AI · July 2026

When you ask a general-purpose chatbot something, the model answers from everything it "knows" — its training, more or less up to date, with no way for you to check where any given claim comes from. NotebookLM, Google's AI research tool, works on a different premise, one the industry calls grounding: you create a notebook, upload your sources (the papers for your review, a clinical guideline, your notes), and from that point on the assistant answers based on that material, with inline citations that point back to the exact passage in the original document. Click the citation and you see the precise excerpt the answer relies on.

For anyone working with scholarly literature, that traceability changes everything: the conversation stops being an act of faith and becomes something you can verify. Under the hood it runs on Gemini, Google's model family; since June 2026, version 3.5 — which gave NotebookLM a major leap forward — lets the chat help you build your source repository, search the web to suggest new documents, and even run code in a secure sandbox to produce analyses and charts. The tool launched in 2023 as a Google Labs experiment and is now a mature product, with a mobile app and a presence in Google Workspace as a Core Service.

One thing worth stating clearly now, and that we'll come back to below, is something Google itself says explicitly: source grounding is a transparency mechanism, not a guarantee of accuracy. NotebookLM can still get things wrong even with the documents right in front of it.

• Each individual source allows up to 500,000 words or 200 MB, whichever comes first. That's a high ceiling — whole books fit. If a file exceeds it, or if the PDF is copy-protected, the import is rejected outright.
• The free plan allows 50 sources per notebook and up to 100 notebooks per user, with 50 chat queries and 3 Audio Overviews a day.
• Paid plans raise those caps: 100 sources per notebook on Google AI Plus, 300 on Pro, and up to 500 or 600 on the Ultra tiers, with daily chat and audio-generation limits scaling up accordingly.

• Audio Overview: a conversational podcast in which two AI-generated hosts discuss your sources. This was the feature that made the tool famous, in September 2024, and since April 2025 it has been available in more than 50 languages, including Spanish. It also has an interactive mode (in beta and, for now, English only) that lets you "join" the conversation and ask the hosts questions live.
• Video Overview: the video equivalent, with narration and visual support, rolled out broadly in July 2025.
• Mind Map: an interactive mind map that organizes the concepts in your sources and lets you navigate them — useful for getting a sense of a corpus's structure before diving into a serious read.
• Study guides, flashcards, and quizzes, built for teaching and self-study.
• Custom Reports: tailored reports where you describe the structure, headings, tone, reading level, and language you want in a prompt, and NotebookLM drafts the document from your sources.
• Infographics, with ten preset styles, including a Scientific one alongside others like Professional, Editorial, or Sketch Note. We'll come back to these in the recipe below.
• Since late 2025 and through 2026, data tables, slide decks, and PDF reports with charts have been added.

Triaging and querying a corpus of articles. The most natural scenario: you've gathered 30 or 40 papers on a topic and need to get your bearings before reading them in depth. You upload the PDFs to a notebook and ask which populations are studied, which outcomes are measured, which articles cover a given subtopic. Every answer arrives with citations to the original passage, so you can jump straight to the relevant paragraph in each article.

Journal club and critical appraisal. To prepare a journal club session, a notebook with the 2 or 3 articles under discussion lets you generate discussion questions, section-by-section summaries, or an Audio Overview attendees can listen to on their way to the hospital.

Comparing methods and results across studies. Questions like "how do the inclusion criteria of these five trials differ?" or "which studies find an effect and which don't?" are ideal territory for grounded chat: they force the model to locate specific passages across several documents and return them side by side — tedious to do by hand and easy to verify through the citations.

Supporting literature reviews, with one clear limit. NotebookLM helps in the exploratory phases of a review: getting familiar with a topic, spotting concepts and vocabulary for search strategies, identifying common threads and contradictions across studies already retrieved. What it doesn't do, and this needs to be clear, is replace the formal process of a systematic review: it doesn't substitute for expert database searching, dual screening, or verified data extraction. It can also get things wrong despite grounding, and any references it generates need to be checked one by one. Industry comparisons agree on how the roles split: tools like Elicit or Consensus discover literature in their own academic indexes; NotebookLM synthesizes the corpus you've already assembled.

Preparing classes and teaching materials. Study guides, quizzes, and flashcards generated from your own materials (your slides, the course readings, a clinical guideline) save hours of prep work. It's no accident that NotebookLM is included free in every edition of Workspace for Education.

Patient materials. A Custom Report instructed to write "for a patient with no health training, at an accessible reading level" from a clinical guideline produces quite reasonable drafts, which a professional should then review, as with any AI output in this space.

Science communication through the podcast. The Spanish-language Audio Overview turns a dry report or a couple of articles into a ten-minute conversation you can share with residents, a board, or a scientific society's audience. We use it in newsletter production, and experience has taught us to check the narration against the original documents before publishing, because the hosts occasionally introduce inaccuracies — Google itself warns as much.

And here's the part we most enjoy telling, because it's homegrown and we actually use it: how to generate a visual abstract or a scientific infographic that follows a specific graphic style, without touching Canva or Figma afterward. To be upfront from the start: NotebookLM already ships with an infographics feature offering ten preset styles, one of them Scientific, and anyone can pick one and request their visual — we didn't invent that part. What we add is one more step so you don't have to depend on those presets and can lock in your own, consistent graphic style. It's not, then, an official feature packaged by Google, but a workflow we've built by combining pieces that already exist. Here's how it works:

• Step 1: extract the style with another AI. We start with a visual reference: a poster or visual abstract we like, or the journal's or institution's brand guidelines. We hand it to a general-purpose assistant (ChatGPT or Claude) and ask it to describe it as a structured YAML or JSON specification: fonts and hierarchies, a color palette with codes, section structure, visual tone. The result is a one-page, machine-readable style guide that captures the essence of the reference design.
• Step 2: paste that guide as a source in NotebookLM. In the notebook that already contains the paper or scientific content, we add the style specification as one more source. A technical detail worth being upfront about: YAML and JSON aren't on the official list of supported file types, so it isn't uploaded as a file — it's pasted in as copied text (which is an official source type). It goes in without issue, and the model treats it as just another document in the notebook.
• Step 3: request the visual. We generate the infographic or visual abstract by asking NotebookLM to follow the style guide among its sources: to use that palette, that typographic hierarchy, and that section structure to synthesize the paper's content. NotebookLM's visual output is the final result itself — there's no need to export the content and lay it out separately in another tool.

We've reached the question that, rightly, worries our audience the most. If you're going to upload unpublished manuscripts, research protocols, or clinical documentation, you need to know exactly what Google does with that content. The good news is that the official policy is fairly clear; the fine print lies in the differences between account types.

The general rule. NotebookLM's official privacy page puts it this way:

In other words, what you upload and what you ask isn't used to train Google's foundation models, with a single exception: voluntary feedback.

The feedback exception. If, on a personal account (the @gmail.com kind, free or paid), you tap thumbs up or thumbs down to rate a response, Google may collect the full context of that interaction (your questions, your uploaded sources, and the model's responses) and submit it to human review by specialized teams. Google disconnects that feedback from your account before reviewers see it, and the reviewed material is kept for up to 3 years, already unlinked from your identity. Google itself explicitly asks that you not include confidential or sensitive information in feedback.

Workspace and Education accounts: the exception disappears. On institutional accounts, protection is stronger, and the official statement deserves to be quoted in full:

No human review, no training, not even when a user submits explicit feedback. Since 2025, NotebookLM has also been a Core Service of Workspace and Workspace for Education, under Google Cloud's contractual data-protection framework — though the classification isn't uniform across every edition (in some it's listed as an additional service), so whoever administers a domain would do well to check the specific edition under contract.

What doesn't exist: a privacy switch. Here's the nuance almost nobody mentions. Unlike the Gemini app, which has an explicit activity setting you can turn off, NotebookLM offers no dedicated setting to opt out of human review on a personal account. The only way to avoid it is behavioral: don't tap the feedback buttons. We think this is a real gap in product transparency, and we think it's worth flagging so our readers know about it.

Our practical recommendation, defensible with the sources in hand: for sensitive documents (unpublished manuscripts, project data, clinical documentation), use an institutional Workspace or Education account, where protection doesn't depend on your behavior, and if you're on a personal account, don't submit feedback on notebooks containing that material. One more caution: the consumer version's privacy page doesn't specify where data is stored geographically; anyone who needs residency guarantees (in the EU, say) will need to look at NotebookLM Enterprise, the Google Cloud variant that does offer that control.

IN ONE SENTENCE

On a personal account, protection depends on your behavior (not tapping feedback); on a Workspace or Education account, there is no human review or training under any circumstances. For sensitive documents, that difference decides which account to use.

We close the analysis with what NotebookLM doesn't guarantee, starting straight from Google:

On the podcasts, the official disclaimer acknowledges that the hosts

and notes that the generated conversation isn't an exhaustive or objective view of the topic, but a reflection of the uploaded sources.

The tech press has documented specific cases of this kind of error: an Audio Overview that attributed to a book details the book never contained and presented a paraphrase as a direct quote, or an answer that recommended unsuitable material for an automotive piece because the sources the user had uploaded were low quality. That second case illustrates the tool's golden rule: garbage in, garbage out. NotebookLM answers from your sources, so the quality of the answer will never exceed the quality of the corpus. Some blogs circulate specific "hallucination rate" figures for NotebookLM versus other chatbots; we don't repeat them here because no study with a verifiable methodology backs them up.

That's where our best practices come from — the same ones we apply in our own work with generative AI:

• Curate the corpus before you ask anything. Select sources with the same rigor you'd apply for a review. A notebook is only as good as what's in it.
• Verify citations, always. Inline citations make verification easy, but not automatic: you still have to click through and read the passage. And every bibliographic reference the tool drafts gets checked against the actual source before you use it.
• Verify derived outputs before sharing them. A podcast or video that's going to leave the private sphere gets checked against the original documents. We do this systematically with the audio and video we produce, and it's not unusual to find inaccuracies that need fixing.
• Remember which phase of the work it covers. Exploration, synthesis of what's already gathered, teaching, and science communication — yes; expert database searching, formal screening, and verified data extraction for a systematic review — no.

Of all the general-purpose AI tools out there, NotebookLM is the one that best fits how scientific documentation actually works, because its design revolves around something non-negotiable in our field: knowing where every claim comes from. Used well, with a curated corpus and systematic verification, it saves many hours and opens up science-communication formats that used to require a full production studio.

If you want to bring it into your hospital, scientific society, or research group with real judgment, it's one of the tools we cover in our training courses, from basic use to advanced workflows like the visual abstract one. And if what you need is information support or guidance for your research with these tools already built into the process, we can help with that too.

• Google: Privacy and Terms of Use in NotebookLM (official privacy page, accessed July 2026).
• Google: NotebookLM FAQ (source limits and deletion).
• Google: Learn about NotebookLM (official error disclaimer).
• Google: Upgrade NotebookLM (plan limits table).
• Google: Add or discover new sources (supported source types).
• Google Workspace Updates: NotebookLM and NotebookLM Plus now a Workspace Core Service (February 2025).
• Google Workspace Updates: NotebookLM and Gemini app Core Services for Education customers (April 2025).
• Google: NotebookLM Audio Overviews (official blog, podcast disclaimer).
• Google: NotebookLM Discover Sources (April 2025).
• Google: NotebookLM introduces curated featured notebooks (July 2025).
• Google Workspace Updates: New ways to customize and interact with your content in NotebookLM (March 2026, infographic styles).
• Google: Google AI subscription plans and Gemini pricing (accessed July 2026).
• Google Cloud: NotebookLM Enterprise overview.
• TechCrunch: NotebookLM's new update will help you build source repository from chat (June 2026, Gemini 3.5 update).
• Kyle Orland, Ars Technica: Fake AI "podcasters" are reviewing my book and it's freaking me out (September 23, 2024).
• XDA Developers: NotebookLM's Audio Overviews sound so human, you'll believe their misinformation.
• The Effortless Academic: Google's NotebookLM updates in 2025 for literature review and study.