privacydeep-dive

NotebookLM Privacy: What Google Does With Your Uploaded Files

NotebookLM lets you upload personal PDFs and documents for AI research. Here's exactly what Google's privacy policy says about your uploaded content.

Google NotebookLM has quietly become one of the most popular tools for working with personal documents. Researchers upload academic papers. Journalists organize source documents. Professionals synthesize reports. Individuals upload personal journals, family histories, medical records — things they’d never post publicly — because the tool is genuinely useful and the interface feels private.

The question worth asking before you upload your first personal document: what does Google actually do with what you upload?

This post answers that question as precisely as the available evidence allows.


What NotebookLM Is, and Why People Upload Personal Documents

NotebookLM is Google’s AI-powered research tool that lets you upload “sources” — PDFs, Google Docs, web pages, audio files, YouTube videos — and then ask questions about them. The AI summarizes, quotes, synthesizes, and generates audio overviews from your uploaded material.

The appeal is obvious. Instead of manually cross-referencing twenty PDFs, you can ask the AI a question and get a synthesized answer with citations.

The documents people upload are often deeply personal:

  • Medical records and test results they want to understand
  • Journals and personal writing projects
  • Family history documents and genealogy research
  • Legal documents they want to parse without paying an attorney by the hour
  • Financial records they’re trying to organize
  • Old letters, diaries, scanned photos with captions

These aren’t public documents. They’re personal archives. And they’re being uploaded to Google’s servers as a matter of routine.


What Google’s Privacy Policy Actually Says

Google Workspace and NotebookLM are covered by Google’s main privacy policy, with additional terms in the Workspace-specific terms and the generative AI additional terms.

Here’s what matters:

Content you upload is not used to train general Google AI models. Google has stated explicitly, in multiple communications, that content uploaded to NotebookLM is not used to train their public AI models. This distinguishes NotebookLM from Google Search, Google Photos with AI features enabled, or other products where data use for AI training is the default.

Your content is processed to deliver the service. When you ask NotebookLM a question about a document, that document’s content is sent to Google’s AI infrastructure for processing. This is true of every AI tool that processes your documents — the question is what happens to it afterward.

Data retention is governed by your account settings. NotebookLM content is stored under your Google account. If you delete a notebook or a source document within NotebookLM, Google’s standard deletion timelines apply — content may persist in backup systems for up to 60 days after deletion under standard Google data retention policies, longer in some enterprise configurations.

Google employees may access content in limited circumstances. The standard Google privacy policy reserves the right to access user content for safety review, legal compliance, and abuse prevention. This is standard across cloud services, but it means your documents are not completely inaccessible to humans.


The Workspace vs. Personal Account Distinction

This matters significantly.

If you use NotebookLM through a personal Google account (your personal Gmail), your content is subject to Google’s consumer privacy policy and the NotebookLM-specific terms.

If you use NotebookLM through a Google Workspace for Business account, the terms are governed by your organization’s Workspace agreement, which typically includes stronger data protection commitments and explicit prohibitions on Google using customer data for product improvement.

If you use NotebookLM through a school’s Google Workspace for Education account, student data protections apply.

Most individuals using NotebookLM casually are using personal Google accounts. This is the configuration where the data terms are most permissive and where the “not used for training” commitment is made as a policy statement rather than a contractual obligation with enforcement teeth.


What “Not Used for Training” Actually Means

The commitment that your NotebookLM uploads aren’t used for AI training is meaningful, but it has limits worth understanding.

It applies to training public AI models. Google’s internal product improvement, debugging, safety review, and quality assurance processes are not the same as training public models. Your document content may be seen by automated systems and potentially by employees under these categories even if it’s not used to train the next version of Gemini.

It applies currently. Privacy policies change. Google has updated its Workspace terms and its AI training policies multiple times in the past few years. The commitment in place today may not be the commitment in place in three years.

“Aggregated and anonymized” data can be used. If Google anonymizes and aggregates usage patterns — which questions users tend to ask, which document types produce which kinds of interactions, frequency and format patterns — this is generally not covered by “no training” commitments because it’s no longer personally identifiable content.


What Other AI Document Tools Do

Comparing to other AI document tools helps contextualize the risk.

ChatGPT (OpenAI). If you use GPT-4 through the consumer ChatGPT interface, OpenAI’s privacy policy (as of early 2026) states that they may use conversations and content for training unless you opt out through account settings or use the API. Opt-out is available but is not the default.

Claude (Anthropic). Anthropic states that conversations and uploaded files are not used to train AI models without explicit opt-in. For API use, no training use applies by default.

Microsoft Copilot. For consumer Copilot, data handling is governed by Microsoft’s consumer privacy policy. For enterprise customers with Copilot for Microsoft 365, stronger data protection terms apply.

The pattern: enterprise/paid tiers generally have stronger, more contractually enforceable data protection commitments. Consumer tiers have more permissive policies that may be harder to enforce if violated.

NotebookLM falls somewhere in the middle — it’s a consumer product with explicit (if policy-level rather than contractual) commitments to not use content for AI training.


Document Categories and Risk Assessment

Not all personal documents carry equal sensitivity. Before uploading anything to NotebookLM or any AI tool, it’s worth a quick risk assessment.

Lower sensitivity (relatively safe to upload):

  • Published books or articles you’re researching
  • Business documents without personal identifiers
  • Technical manuals, academic papers
  • News articles you want to synthesize

Medium sensitivity (think before uploading):

  • Work documents with client information
  • Personal writing projects without identifying information
  • General financial overviews without account numbers

Higher sensitivity (upload with caution or don’t upload):

  • Medical records with diagnosis information
  • Legal documents with case-sensitive details
  • Financial records with account numbers or SSN
  • Personal journals with identifying information about others
  • Family documents with third parties’ information

The useful heuristic: would you be comfortable if a smart Google employee who respects privacy could see this document in the course of their job? If yes, the risk is relatively low. If no, reconsider.


Technical Privacy Considerations

Data transmission. Documents are uploaded over HTTPS — standard encrypted transmission. This protects them from interception in transit.

At-rest encryption. Google encrypts data at rest using AES-256. This protects against unauthorized access to storage infrastructure but does not prevent Google itself from accessing the content.

Access controls. NotebookLM notebooks are private by default. Sharing a notebook gives the recipient access to the sources and the notebook’s interaction history.

Cross-product data use. NotebookLM is tied to your Google account. The fact that you uploaded a specific document to NotebookLM becomes part of your Google account activity. Whether this feeds into Google’s broader understanding of your interests and behaviors is not explicitly addressed in NotebookLM-specific documentation.


The Alternatives Worth Considering

If you need AI-assisted document research on sensitive personal content, there are options with different privacy tradeoffs:

Local processing. Running a local language model (Ollama, LM Studio) on your own machine with a tool like Open WebUI keeps document processing entirely offline. Performance depends on your hardware and the model, but for many document tasks it’s sufficient.

API with privacy terms. Using the Anthropic Claude API directly, or OpenAI’s API with training opt-out enabled, processes your documents without storing them for training purposes — and with API-grade data handling terms rather than consumer terms.

On-device processing. Apple Intelligence’s on-device processing (for documents processed entirely on-device) and similar on-device AI features don’t transmit your content to external servers.

None of these are as convenient as NotebookLM’s interface. That’s the privacy-convenience tradeoff.


The Practical Recommendation

NotebookLM’s “not used for training” commitment makes it meaningfully safer than some alternatives for working with personal documents. It is not the same as your documents being private from Google.

For most informational documents — research material, public records, published texts — the risk profile is acceptable for most people.

For genuinely sensitive personal documents — medical records, legal filings, financial statements, personal journals — the safer approach is either a tool with stricter contractual protections, a privacy-focused alternative, or local processing.

The rule that applies everywhere applies here: once you upload personal data to a cloud service, you’ve shared it with that service. Understanding exactly what that means before you click upload is worth the few minutes it takes.

Your memories deserve better than an ad platform.

Try daftei free →
← All posts