Understanding your tools · Card 02
Understanding
NotebookLM
What makes it different from other AI tools, what it can and cannot do, and how to use it ethically in qualitative research.
What makes NotebookLM different
Document-grounded
Answers come from your sources
You upload your own materials — PDFs, transcripts, notes. Every response is cited from those specific documents, with clickable passages. Unlike general chatbots, it cannot draw on the wider web.
Oxford-licensed
Not the same as a free account
Via Oxford's Google Workspace for Education: your data is not used to train AI models, not human-reviewed by Google, and stays within the University's institutional domain.
⚠️
The Oxford licence changes the data protection picture significantly — but it does not substitute for explicit participant consent. Both protections are needed together for research involving human participants.
What it can and cannot do
Appropriate uses
- Synthesise across up to 50 uploaded documents simultaneously
- Generate structured summaries to support early-stage analysis
- Identify recurring language patterns across a set of sources
- Build a searchable, queryable notebook from fieldnotes or memos
- Create an audio overview of your literature for a different mode of engagement
- Surface patterns and connections for you to interrogate
Important limits
- ~13% hallucination rate even with source-grounding — always verify outputs
- Cannot detect tone, silence, embodied meaning, or power dynamics
- Cannot act as primary interpreter — your judgement cannot be delegated
- Model version may change without notice, limiting reproducibility
- Broad prompts can cause characterisation that drifts beyond the source material
- Not appropriate for highly sensitive data without explicit ethics review
If processing participant data — consent is non-negotiable
Four things your consent process must cover
1
Name the tool. Your Participant Information Sheet must name NotebookLM, state it is cloud-based, and explain what happens to their data — including the Oxford licence protections.
2
Explain the data pathway. Participants must know their data travels to a third-party server. Plain language; no jargon.
3
Enable genuine refusal. Participants must be able to decline AI processing while still taking part. This must be a real choice, not a formality.
4
Tie withdrawal to AI outputs. If a participant withdraws, remove not just their raw data but any AI-generated summaries or codes derived from it.
Choosing the right tool for your data
Tier 1
Truly local AI (LM Studio, Ollama) — for highly sensitive data: detailed trauma accounts, clinical data, politically sensitive material, data under NDA. Data never leaves your device.
Tier 2
Oxford-licensed NotebookLM — where explicit participant consent has been obtained, data is appropriately anonymised, and ethics approval covers AI use.
Tier 3
Consumer cloud AI (free accounts) — for non-personal data only: published literature, your own notes, draft text. Consumer terms allow data use for training. Not appropriate for research data.
Analogy
Think of NotebookLM as a very thorough research assistant who has read all your uploaded sources and can retrieve connections across them — but who still needs you to decide what those connections mean. The assistant surfaces; you interpret. The interpretive responsibility is always yours.