Digital Digging × AI Training for TV 2 Journalists
20 Exercises for NotebookLM
Your documents, your questions, your footnotes
April 2026 | 20 Exercises | NotebookLM
Exercises 0 -- 1
Start Together
The Summary Comparison
What you will learn
That AI does not give the same answer twice. Everyone at your table uploads the same PDF, asks the same question, and gets a different summary. This is the most important lesson before you start working with any AI tool.
What to watch for
AI is stochastic — it generates text probabilistically, not deterministically. The same input never produces the same output. This means no AI summary is “the” summary. It is “a” summary. That distinction matters for every piece of journalism you will ever produce with AI assistance.
TV 2 connection: Everyone — this is the foundation exercise. Every participant needs to see with their own eyes that AI output is variable before they start relying on it.
The Prewash — Table Exercise
What you will learn
The Prewash Method: instead of writing your own prompt, you let the AI design the analysis. Two commands, that is it. Then you compare results at your table to see how the same method produces different analytical frameworks for different people.
1. Scope
"What exactly did Rigsrevisionen audit, what time period does it cover, and which municipalities or ministries are named?"Why: prevents misreporting “all municipalities” when only a sample was studied.
2. Method
"What method did the auditors use — sample size, case selection, statistical vs. legal compliance review — and what are its acknowledged limitations?"Why: a neutral methodology question is the single best defence against amplifying a flawed finding.
3. Definitions
"Which specific legal requirements were tested, and how is 'non-compliance' defined in this report?"Why: forces the document to declare its yardstick before any number is reported as a scandal.
4. Counter-voice
"What does the Ministry of Social Affairs and Housing say in its formal response, and where does it agree or disagree with Rigsrevisionen's conclusions?"Why: every Rigsrevisionen report includes a ministerial reply — the disagreement is often the real story.
5. Historical baseline
"Which findings are new in this report, and which are repeats of issues already raised in the 2016 audit that this report follows up on?"Why: distinguishes “newly discovered problem” from “decade-old problem the state still hasn’t fixed” — two very different headlines.
What to watch for
Compare the Prewash result with the simple summary from Exercise 0. The Prewash should produce a significantly deeper analysis — not just what the document says, but what it implies, what it leaves out, and what questions it raises. That is the difference between summarising and analysing.
TV 2 connection: Everyone — this is the core technique you will use for all document work with AI. Two commands. Nine words.
Exercises 2 -- 9
Beginner
Add the 2016 Audit — Design Your Own Comparison
What you will learn
There are two Rigsrevisionen reports on the same problem — the 2016 audit and the 2026 follow-up. Same auditor, same topic, ten years apart. The journalistic question writes itself: what changed, what did not, and what was quietly dropped? Your job in this exercise is not to be told the comparison — it is to design it yourself using NotebookLM as your thinking partner.
What to watch for
When AI compares two documents on the same topic, it tends to produce one of three failure modes: (1) collapse — treating both as one document and losing track of which fact came from where, (2) false symmetry — inventing a 2016 equivalent for every 2026 finding even when none exists, and (3) narrative drift — turning what should be a side-by-side comparison into a tidy story arc (“things have improved” or “things have gotten worse”) that may not reflect the actual evidence. The footnote-verification step in (6) is your only defence.
TV 2 connection: Lars — the accountability beat. Two audits, ten years apart, one question: what did the state promise to fix, and is it fixed? You are not just running a comparison — you are learning to design one, which is the skill that scales to every future investigation.
13 Documents, One Question
What you will learn
How to load multiple documents into a single NotebookLM notebook and ask one question that forces the AI to cross-reference all of them. This is where NotebookLM becomes a genuine research tool — it can read 13 PDFs simultaneously and find patterns no human would spot in a reasonable time.
Contents:
• Politikredse_2024.pdf — annual breakdown by police district
TV 2 connection: cross-referencing 13 documents in one notebook is the kind of synthesis that would take a journalist a full day by hand — and a footnote-checked AI run in minutes.
Fyn Focus — 2024 Only
What you will learn
How to narrow a multi-document notebook down to a single source and pull out beat-relevant numbers. The skill: writing one sharp prompt that returns a usable answer — or a clear “the data is not in this document.” Both outcomes are stories.
Politikredse_2024.pdf for this round (or de-select the other sources in the source panel).Politikredse_2024.pdf source, find out for Fyns Politi: in which month of 2024 were there the most accidents, in which the fewest, and what was the spread?TV 2 connection: Fyn is in Fyns Politi. This exercise lets every journalist in the room find out whether their own district is getting safer or more dangerous — and in which months. Unfortunately we have to wait till the 2025 report comes out to do that.
The Four-Year Comparison
What you will learn
How NotebookLM handles two related documents at once — comparing what changed, what stayed, and what quietly disappeared between two government reports four years apart. This is the bread-and-butter of policy journalism: tracking promises against follow-through.
What to watch for
NotebookLM is good at surfacing differences, but “absent in 2023” can mean three different things: (1) genuinely abandoned, (2) renamed or merged into another initiative, (3) achieved and therefore not re-stated. The model often cannot distinguish these. Your job as a journalist is to take the dropped-reform list as a starting point for reporting, not as a finding.
TV 2 connection: Lars — the government accountability beat. Two documents, four years apart, one prompt: instant story leads about what the state stopped talking about.
The Audio Summary
What you will learn
How to use NotebookLM's Audio Overview feature to generate a podcast-style summary of your documents -- and how to verify what the audio claims against the actual source material.
What to watch for
The Audio Overview is designed to be engaging, which means it can oversimplify or dramatize findings. The "hosts" sometimes make connections or draw conclusions that go slightly beyond what the document actually states. The audio is a starting point for understanding, not a substitute for reading. Always verify specific claims back to the footnoted source.
TV 2 connection: Mikkel Fruerboel Secher -- using audio as an alternative pathway into document analysis, particularly useful for documents in languages you are translating from.
The Footnote Verification Drill
What you will learn
How reliable NotebookLM's footnotes actually are -- by systematically testing ten questions and scoring the accuracy of every citation. This builds calibrated trust: knowing when to rely on the tool and when to double-check.
What to watch for
NotebookLM typically scores very high on factual extraction (numbers, names, dates) but can stumble on interpretive questions where the answer requires synthesizing multiple passages. Pay special attention to answers where the footnote points to the right section but the paraphrase subtly shifts the meaning. That is the most dangerous type of error.
TV 2 connection: Lars Apel -- directly addressing trust issues with AI by building an evidence-based understanding of when NotebookLM's footnotes can be relied on.
Upload and Compare Two Documents
What you will learn
How to use NotebookLM to compare two versions of the same document -- draft vs. final, old regulation vs. new, original statement vs. revised. This is where NotebookLM's multi-document capability becomes genuinely powerful.
What to watch for
NotebookLM may miss small but important changes (a single word replaced, a sentence deleted) while focusing on large structural differences. For a thorough comparison, follow up with: "Are there any small changes -- single words or phrases -- that might be significant?" Also watch for changes NotebookLM correctly identifies but misinterprets the significance of.
TV 2 connection: Pelle Lykkebo Mork -- tracking policy evolution and spotting what changed between draft and final versions of legislation or regulation.
The Quote Finder
What you will learn
How to extract all direct quotes from a specific person in a long transcript or document -- and verify each one against the source. Essential for working with interview transcripts, hearing protocols, and meeting minutes.
What to watch for
NotebookLM sometimes paraphrases when you asked for direct quotes. Always verify word-for-word accuracy before using any quote in your reporting. Also, it may miss quotes where the person is referred to by a different name or title elsewhere in the document. Try: "Also search for quotes attributed to [title/role] that might be the same person."
TV 2 connection: Marie Moller Munksgaard -- expert research and pulling precise quotes from lengthy source material for context-building.
Notebook Organization
What you will learn
How to organize a NotebookLM notebook as a proper research environment -- multiple sources, saved Notes, and a structure that makes your analysis retrievable and buildable.
What to watch for
NotebookLM's Notes are not automatically organized -- you need a naming convention. Without one, a notebook with 20+ Notes becomes unusable quickly. Decide early: will you organize by theme, by date, or by document? Consistent naming will matter when you start building permanent beat archives in the advanced exercises.
TV 2 connection: Anne Fuglsang Borg -- building organized, sortable research environments from multiple sources of data.
Exercises 10 -- 16
Intermediate
Questions You Can't Get from the Table of Contents
What you will learn
How to ask questions that go beyond summarization -- the kind of questions that require reading the entire document carefully and making connections between different sections. This is where NotebookLM becomes genuinely useful for journalism.
What to watch for
NotebookLM excels at these cross-sectional queries because it can "read" the entire document simultaneously. However, it may flag routine legal exceptions as "contradictions" or interpret standard boilerplate as "surprising." Your editorial judgment determines whether a finding is genuinely newsworthy or just technically interesting.
TV 2 connection: Mathias Overgaard -- going beyond summaries to extract the specific details that matter for OSINT work and error-free reporting.
Multi-Document Comparison
What you will learn
How to track changes and contradictions across multiple years of the same type of document -- annual reports, budget proposals, performance reviews. This is where NotebookLM's multi-source capability reaches its full potential.
What to watch for
NotebookLM may confuse changes in reporting methodology with changes in actual performance. A number going down might mean performance declined, or it might mean they changed how they count. Always ask: "Did the methodology change between years?" Also, NotebookLM may not notice when something present in an earlier report simply disappears from a later one -- omission is harder to detect than contradiction.
TV 2 connection: Mathias Overgaard -- tracking organizational claims over time to identify where the narrative shifts or where promises quietly disappear.
Verdict Analysis
What you will learn
How to use NotebookLM to systematically analyze a court verdict -- extracting key findings, decisive evidence, and both parties' arguments in a structured, footnoted format.
What to watch for
Legal language is precise and NotebookLM's paraphrasing may inadvertently shift the meaning. A court "finding" is not the same as a court "opinion" or "suggestion." Always verify that NotebookLM's summary uses the correct legal weight. Also, verdicts often reference previous rulings -- NotebookLM cannot access those external references, so it may miss important legal context.
TV 2 connection: Jakob Hohlmann Villumsen -- quickly analyzing verdict summaries and extracting the legally decisive elements for reporting.
Cross-Referencing Press Release vs. Source
What you will learn
How to catch spin by comparing a press release against the underlying report or study it claims to represent. NotebookLM's footnotes make it possible to verify exactly where a press release accurately represents its source -- and where it distorts or omits.
What to watch for
NotebookLM may flag simplification as misrepresentation. A press release is supposed to simplify -- the question is whether the simplification changes the meaning. Focus on cases where the press release claims something the report does not support, uses numbers differently, or omits findings that contradict the press release's narrative. Those are the story leads.
TV 2 connection: Peter Moller -- verification of scientific sources and press releases against the actual underlying research.
The Regulatory Tracker
What you will learn
How to systematically compare current and proposed regulations to identify what actually changes, who is affected, and how enforcement shifts. This is the regulatory change analysis workflow.
What to watch for
Regulatory language is often deliberately vague. NotebookLM may present a change as clear-cut when the actual legal effect is ambiguous. Pay attention to phrases like "may" vs. "shall", "reasonable" without definition, and new exemption clauses that could undermine the regulation's stated purpose. These nuances require human judgment.
TV 2 connection: Pelle Lykkebo Mork -- policy analysis and tracking regulatory changes that affect newsroom operations and public interest.
Build a Source Dossier
What you will learn
How to use NotebookLM to build a comprehensive dossier on a source by uploading everything publicly available about them and letting the tool find patterns and contradictions across documents.
What to watch for
NotebookLM can only work with what you upload. If a critical piece of public information is missing from your notebook, the dossier will have a blind spot. Also, "contradictions" may reflect legitimate changes of opinion over time -- always check the dates and context. A person who changed their view after new evidence emerged is not contradicting themselves in a journalistically meaningful way.
TV 2 connection: Nanna and Jakob -- building source profiles for interview preparation and background research on key figures.
The Meeting Archive
What you will learn
How to analyze a series of meeting minutes to detect recurring topics, broken promises, and items that were raised but never resolved. Turning bureaucratic paper trails into story leads.
What to watch for
Meeting minutes vary enormously in quality and detail. Some minutes summarize decisions; others record discussions verbatim. NotebookLM works best with detailed minutes. With sparse minutes, it may infer connections that the text does not actually support. Also, a topic "disappearing" from the agenda might mean it was resolved informally -- not every disappearance is a story.
TV 2 connection: Sanne Lau Pedersen -- monitoring committees and councils over time to catch patterns that individual meeting coverage would miss.
Exercises 17 -- 22
Advanced
Stress-Testing the Source Limit
What you will learn
How to stress-test NotebookLM with a large collection of documents on the same topic -- discovering what it handles well at scale and where it starts to lose coherence. The title is aspirational; start with 10+ and push toward the limits.
What to watch for
NotebookLM currently supports ~50 sources per notebook. As the number grows, it may start favoring certain documents over others in its answers -- typically the ones uploaded most recently or the ones most textually similar to your question. Check whether the footnotes draw from a wide range of your sources or cluster around a few. If answers become shallow, try being more specific in your questions.
TV 2 connection: Sanne Lau Pedersen -- monitoring workflows that accumulate large document collections over time, testing whether NotebookLM can serve as a long-term research archive.
Building Your Beat Archive
What you will learn
How to create and maintain a permanent NotebookLM notebook for your beat -- a living research archive that grows with every document you encounter, eventually revealing stories that no single document could tell.
What to watch for
The value of a beat archive is cumulative -- it becomes more useful as it grows. But it also requires curation: not every document is worth uploading. Be selective. Also, NotebookLM's ~50 source limit means you may need to rotate older documents out or create sub-notebooks by theme. The tool is a research assistant, not a filing system.
TV 2 connection: Franziska Weiss Lauritzen -- building a permanent criminal journalism research archive that accumulates intelligence across cases and time.
Criminal Network Documentation
What you will learn
How to use NotebookLM to map connections between individuals across multiple court documents, news articles, and public records -- building a documented network analysis from text sources.
What to watch for
NotebookLM may confuse people with similar names or roles. It cannot verify identity -- if "Lars Jensen" appears in two different documents, it may assume it is the same person when it is not. Always verify identity claims independently. Also, connections that exist in the documents may not represent real-world connections -- co-mention is not the same as collaboration.
TV 2 connection: Franziska Weiss Lauritzen -- criminal journalism and OSINT, mapping networks from publicly available case documentation.
The Methodology Reviewer
What you will learn
How to use NotebookLM to critically analyze the methodology of academic studies -- with every critique anchored to specific passages in the paper. The footnotes ensure you can verify whether NotebookLM's critique is grounded or speculative.
What to watch for
NotebookLM is grounded in the uploaded document, so it can only critique what the paper describes. If the methodology section is vague, NotebookLM's analysis will be vague too. It cannot compare the methodology to other studies in the field unless you enable web access -- but doing so can pollute your analysis with outside information. For domain-specific methodological issues, you still need an expert source. Use NotebookLM for the first pass, then consult a specialist.
TV 2 connection: Peter Moller -- evaluating scientific sources with rigor, building the critical analysis skills needed to assess studies cited in press releases and policy documents.
The Timeline Reconstructor
What you will learn
How to reconstruct a chronological timeline of events from multiple documents published at different dates -- and identify where sources disagree on when things happened. Timeline discrepancies are often the most revealing finding in investigative work.
What to watch for
NotebookLM may confuse "when a document was published" with "when an event occurred." A report published in March 2025 about events in November 2024 should be placed at November 2024 in the timeline, not March 2025. Verify that the timeline reflects event dates, not publication dates. Also, timeline gaps are as important as timeline conflicts -- what period does no document cover?
TV 2 connection: Pelle Lykkebo Mork -- breaking news reconstruction, building forensic timelines from multiple sources to identify where the official story does not hold.
Design Your NotebookLM Workflow
What you will learn
How to design a complete, personalized NotebookLM workflow -- from which documents to upload routinely, to what questions to ask, to how to organize your findings. The goal: a system you will actually use after training ends.
What to watch for
The most common failure mode for AI workflows: they are too complex. If your workflow has more than 5 steps for a routine document, you will abandon it within a week. Start minimal. A workflow you actually use every day is infinitely more valuable than a comprehensive one you use once. Also, resist the temptation to use NotebookLM for everything -- some documents are faster to just read.
TV 2 connection: Emil Gjerding Nielson and Lars Apel -- designing sustainable editorial AI workflows that become part of the daily routine, not a training exercise that ends when the course does.
When to use NotebookLM -- and when not to
Use NotebookLM when
- +You have specific documents and need answers grounded in those documents
- +You need to verify that every AI claim traces back to a source passage
- +You need to compare, cross-reference, or find contradictions across multiple documents
- +You want an audio summary to quickly understand a long document
- +You need to build a searchable research archive for an ongoing investigation
- +You want the AI to stay strictly within the evidence (no speculation from outside sources)
Do NOT use NotebookLM when
- -You need internet information but want to keep your analysis clean -- NotebookLM's web access can pollute your research
- -You need to analyze images, video, or non-text content
- -You need writing assistance, rewriting, translation, or creative work
- -Your question requires information that is NOT in the uploaded documents
NotebookLM's superpower is also its limitation: it only knows what you give it. This makes it the most trustworthy AI tool for document analysis (every claim has a footnote) but the worst tool for anything requiring outside information. Use it where grounding matters most. Use ChatGPT, Claude, Gemini, or Perplexity for everything else.