Google NotebookLM Under the EU AI Act: Minimal Risk, Real Confidentiality Duties

For the deploying organisation, Google NotebookLM is minimal risk. The obligations that actually matter are not in the AI Act's high-risk regime at all — they sit in GDPR, confidentiality, and what you do with the outputs.

NotebookLM is Google's research and note-making assistant. You upload your own sources — PDFs, Google Docs, web pages, pasted text — and it summarises them, answers questions grounded only in those sources with inline citations, and generates outlines, study guides and briefings. Its defining trait is that it stays inside what you give it: it does not classify people, score applicants, or make decisions about anyone's rights. As internal knowledge work, it sits at minimal risk under Regulation (EU) 2024/1689. The work to do is real, but it is about the documents you feed it and where the output goes — not about conformity assessments.

This page sets the minimal-risk baseline, names the GDPR and confidentiality duties that do bite, flags the one Article 50 edge case, and walks a worked example.

---

The Baseline: Minimal Risk

Summarising your own documents, asking questions about a research corpus, and generating internal briefings are not regulated high-risk activities. NotebookLM does not touch any Annex III use case in its standard internal use, so there is no registration, no conformity assessment, no Article 11 technical documentation, no FRIA.

The one duty already live for every organisation is Article 4 AI literacy (in force since 2 February 2025). Staff relying on NotebookLM should understand, proportionate to their role, what it does and where it can be wrong. Source-grounding reduces but does not eliminate error — a summary can still misrepresent or over-generalise a source, and citations should be spot-checked. That understanding is the Article 4 obligation in practice.

NotebookLM runs on Google's underlying general-purpose AI models. Those GPAI model providers carry Chapter V duties (Article 53) for the models; that is upstream of you, not a deployer obligation.

---

What Actually Bites: GDPR and Confidentiality on Uploaded Documents

The genuine risk with NotebookLM is not its classification — it is what you upload into it. Every document you add is data leaving your direct control and entering a third-party service, which raises two distinct duties.

GDPR. If the documents contain personal data — client files, employee records, meeting notes naming people — uploading them is processing under the GDPR. You need a lawful basis (GDPR Article 6), an appropriate data processing agreement with Google for the version you use, clarity on data location and retention, and a defensible position on data minimisation. Special-category data (health, ethnicity, religion, trade-union membership) is governed by GDPR Article 9 and needs explicit consent or another specific legal ground before it goes anywhere near an upload. Consumer and enterprise tiers differ on data handling and training commitments — confirm which tier you are on and what its terms say before uploading anything sensitive.

Confidentiality. Independently of GDPR, uploading commercially confidential material — unsigned contracts, board papers, source code, trade secrets, privileged legal documents — can breach NDAs, client engagement terms, or your own information-security policy. The fastest real-world failure mode here is not a regulator; it is an employee pasting a confidential client document into a notebook the client never agreed to. Set an acceptable-use policy that says, in plain terms, which categories of document may and may not be uploaded.

---

The One AI Act Edge Case: Publishing the Output

Internal use stays minimal. But if you publish NotebookLM's generated output externally — a briefing, a report, an AI-generated audio overview pushed to the public — you may step into Article 50(2), which requires AI-generated synthetic content to be marked as such in a machine-readable way (with narrow exceptions for clearly creative or assistive editing). Content-marking applies from 2 December 2026. This is a disclosure duty, not a conformity assessment, and it only arises when the output is for an audience beyond your own walls. Used as an internal research assistant, NotebookLM does not trigger it.

---

The Deadlines That Apply

Because NotebookLM is minimal risk in normal use, the high-risk 2 December 2027 date is not your concern unless you somehow wire its outputs into a regulated high-risk decision. What is live: Article 4 literacy (since 2 February 2025), all GDPR duties (now), and — only if you publish output externally — Article 50(2) content-marking from 2 December 2026. The Article 5 prohibitions are in force too, though ordinary research use does not touch them.

---

Worked Example: Meridian Advisory OÜ

Meridian Advisory OÜ, an Estonian consultancy with ~40 employees and ~€6M turnover, adopts NotebookLM across the firm. The compliance picture turns entirely on document hygiene, not classification:

• Summarising public regulatory texts and the firm's own published thought-leadership to build internal briefings — minimal risk. Article 4 literacy for the consultants; nothing else required.
• A consultant uploads a client's confidential strategy deck to query it faster — this is the real exposure: it processes the client's data in a third-party tool and may breach the engagement's confidentiality terms, independent of the AI Act. Meridian's response is a document acceptable-use policy (no client-confidential or personal data without sign-off and a DPA) plus Article 4 training, not a conformity assessment.
• Publishing a NotebookLM-generated market briefing to clients and on its website — Meridian marks it as AI-generated to stay clean under Article 50(2) (content-marking from 2 December 2026).

Net result: minimal-risk tool, zero high-risk obligations — but a genuine GDPR-and-confidentiality programme around what gets uploaded, which is where the actual risk lives.

---

How Confir Helps

Register NotebookLM in your AI inventory and Confir's rule-based engine confirms the minimal-risk classification, then routes you to the duties that actually apply — the Article 4 literacy record, the GDPR cross-map for documents containing personal data, and the Article 50(2) flag if you publish output externally. The engine is deterministic: the same intake produces the same finding, and the rule that fired is shown in plain language, so a reviewer can see exactly why a research tool lands at minimal risk and where its real obligations sit.

---

Related guides

• minimal-risk AI explained
• Article 4 AI literacy requirements
• GDPR and EU AI Act crosswalk
• Article 50 transparency obligations
• data leakage risks in AI tools
• how to classify AI by use
• AI system inventory management