Digital Digging × TV 2 — Odense

AI Manual for Journalists

A practical guide — from source research to document analysis

Henk van Ess digitaldigging.org 12 April 2026

Disclaimer

Yes, AI can hallucinate. Yes, it is “stochastic parrot” technology. Yes, there are ethical questions. This guide teaches you to use it responsibly, or to recognize when colleagues or sources are using it poorly.

AI is a mirror that reflects what you ask. It has no objective truth. Your words determine the outcome entirely.

The golden rule

AI points you to sources, clarifies them or compares them.
Never use AI itself as the primary source.

Overview

Table of Contents

Chapter 1

The Framing Experiment

Your own intake forms were analyzed by AI. The same 20 answers, three different instructions — three completely different outcomes. Recognizable? It is exactly what you do daily with sources.

Critically framed

Instruction: “Be honest, direct. Don’t hold back.”

Jakob “still browsing Facebook manually”
Sanne’s “automation attempts failed”
Lars “finds Google faster”

Neutrally framed

Instruction: “Analyze in a business-like manner, without value judgment.”

Jakob: “Score 3, Track Beginner”
Sanne: “Score 5, attempted Python monitoring”
Lars: “Score 7, broadest tool experience”

Positively framed

Instruction: “Be motivating and constructive.”

Jakob’s “most specific investigative use case”
Sanne is “most technically ambitious”
Lars is “pushing real boundaries”

The conclusion

The same 20 colleagues, the same scores, the same answers. Three completely different stories. AI has no objective truth. It reflects back what you ask. The words you choose determine the story AI tells.

What this means for your work

✗ NEVER use subjective terms:

Important

Interesting

Notable

Problematic

Controversial

✓ DO use objective, measurable terms:

“Which claims are not supported by sources?”

“Which figures deviate from official data?”

“How many times is the opposing party quoted?”

“Which definitions are vaguely formulated?”

“Where is a time reference missing?”

Chapter 2

What is AI? (and what not)

Google

“Here are 30 websites about OSINT tools. Figure it out yourself.”

AI

“I analyze the sources and give you the answer with references.”

What AI CAN do

Search through reports — 289 pages → direct answer to “what are the key findings?”
Translate academic jargon — “difference-in-differences design” → plain language
Find sources and experts — tools with internet (Perplexity, Gemini) provide current sources with links; offline models (ChatGPT, Claude) can suggest where to look but cannot provide live links
Ask questions you forget to ask — methodological weaknesses, omitted definitions
Transcribe interviews — with clickable, searchable transcripts

What AI CANNOT do

Guarantee facts — can hallucinate (invent sources)
Replace your editorial judgment — doesn’t know what is newsworthy
Write originally — recombines language patterns, invents nothing new
Understand context — it misses what you as a journalist know about your sources
Guard nuance — it likes to produce smooth text, even at the cost of precision

Chapter 3

The 5 Most Important AI Tools

Not every AI model is the same. Here you learn WHEN to use which. Each model has its own personality — comparable to restaurants.

Tool 1: Google NotebookLM

Document Analyst

Cooking with your own ingredients

“You upload your own documents, and only those are analyzed. Nothing added, nothing beside. Doubt whether AI understood it correctly? Check the footnotes that link directly to the source. Transparent and precise — but you have to do the grocery shopping yourself.”

What does it do?

Analyzes only documents that YOU upload
Always provides source citations with clickable numbers (1 2 3)
Almost never hallucinates (works only with your sources)
Up to 300 documents per notebook
Can compare multiple documents with each other

Who among you needs this?

Mads Oxlund Petersen — document analysis research • Mathias — searching through large reports while double-checking everything • Pelle — breaking news verification documents

Practice example

Upload a Danish government policy document. Do a Prewash: “Give me a prompt to analyse (subject).” But you can also ask targeted questions that you cannot derive from the table of contents:

Which organizations, companies or platforms outside the authors are mentioned by name, and in what context?
Which payments, rights or obligations are tied to a condition that lies outside this document itself?
Where does an exception to a prohibition or restriction appear that is stated elsewhere in the document?
Which formulas, ratios or thresholds limit a decision that could otherwise be made unilaterally?
Where does this document make distinctions between the people it applies to?

These are questions whose answers cannot be derived from the title or table of contents — exactly the kind of questions where AI adds value. Each answer contains footnotes that link directly to the relevant passages.

Website: notebooklm.google.com | Free

Tool 2: Perplexity

Internet Researcher

Home delivery — but check the order

“Home delivery during a snowstorm: you have to order what you want yourself and the free version can really take a while. Finding experts and current affairs often requires multiple questions if you don’t have an account.”

What does it do?

Searches the live internet
Provides sources with every answer
Finds experts and contact details
Combines multiple sources

⚠ WARNING — Mathias warns correctly

Mathias wrote in his intake: “I often spend considerable time double-checking.” That’s the right instinct. The free version makes errors. Perplexity tends toward confirmation — if many sources claim something, Perplexity confirms it even when it’s wrong. Always check the source links.

Who among you needs this?

Peter — source research (but beware: AI invents sources) • Franziska — criminal network OSINT research • Jakob — finding people for FoI and verdict research

Practice example

“Which European countries changed OSINT legislation in the past year? Give sources.”

Result: current sources with links to trade publications, policy databases and legal journals. In 2 minutes what would otherwise cost an hour of Googling. But: check every link.

Website: perplexity.ai | Free basic, Pro €20/month

Tool 3: ChatGPT

Fast & Broad

The McDonald’s — comfortable but careful

“Comfortable, predictable, and always available — but not the most refined. Even if you write ‘But that answer is wrong!’ you get responses like ‘Great point, let’s look at this again.’”

⚠ WARNING: sycophancy

ChatGPT is the MOST flattering model. It agrees with you, even when you are wrong. It says “What a great point!” even when your point is not great. This is called sycophancy. Dangerous for journalists seeking confirmation for a thesis.

When YES:

• Translation tasks

• Headline brainstorming

• Standard emails

• Quick fact-check starting point

When NO:

• Deep document analysis

• Source research you rely on

• When you need to be sure of facts

• When you need honest feedback

Who among you already uses this?

Most of the group already uses ChatGPT: Sebastian — basic ChatGPT • Joachim — regular ChatGPT user • Nanna — ChatGPT/CoPilot basics • Mikkel — research and brainstorming • Mads Buur Bach — Gemini Q&A alongside it. Roughly half the group already uses it daily.

Website: chatgpt.com | Free basic, Plus €20/month

Tool 4: Claude

Analytical & Honest

The gastronomic restaurant — business-like and precise

“Business-like, structured, emphasis on quality and accuracy. Serves a thoughtful dish without frills. It is also the only AI that sometimes refuses to execute tasks, like a chef who absolutely will not cook a steak well done.”

When do you use it?

Analyzing reports where reliability is crucial
Dissecting legal documents (legislation, contracts, verdicts)
Letting it WRITE prompts (best prompt-writer of all models)
When ChatGPT gets too enthusiastic — Claude gives honest answers
Rewriting texts in a specific style

Who among you needs this?

Lars — considering switching from CoPilot to Claude • Pelle — OSINT methodology where precision matters • Emil — editorial AI tool feedback • Christian — who is skeptical about AI knowledge and needs honest analysis

Website: claude.ai | Free basic, Pro €20/month

Tool 5: Google Gemini

Deep Research

All you can eat — broad and solid

“The all you can eat restaurant: it serves all kinds of flavors, it is adequate — not spectacularly good or spectacularly bad. With Deep Research it creates a full research report with sources in 15 minutes. With @synthid you can check whether an image was created with Google AI.”

What does it do?

Deep Research: creates its own research plan, visits sources, produces a report (15 min)
Excellent at image recognition and image analysis
Can analyze videos
Daily alerts for news monitoring (smarter than Google Alerts)
@synthid: check whether an image is AI-generated

Who among you needs this?

Mads Buur Bach — video verification and editing • Christian — image verification • Franziska — criminal network OSINT • Lars — news monitoring (agents for news scanning)

Unique feature: Deep Research

Gemini creates its own research plan, asks “Do you want this plan?”, visits sources, and produces a report with source citations. You can watch live which sites it visits. Takes about 15 minutes. Ideal for background research.

Website: gemini.google.com | Free basic, Pro €25/month

Chapter 4

The Prewash Method

This is the most important technique from the entire training. Instead of inventing perfect prompts yourself, you let AI do the work. AI knows better than you which questions it can ask a document — it knows thousands of analytical frameworks that you don’t.

Why writing prompts yourself doesn’t work

You don’t know what AI CAN ask a document

You forget methodological weaknesses, omitted definitions, implicit assumptions

You unconsciously use subjective or leading terms

You waste time perfecting while AI can do it better

The method: step by step

Before you start: upload your document (report, legislation, research paper, press release, annual report, government document).

1

Type:

“Give me a prompt to analyse (subject)”

AI designs an analytical framework with questions you hadn’t thought of: methodological weaknesses, definition issues, structural patterns, comparisons with external data.

2

Type:

“Execute prompt”

AI now analyzes with its own prompt. The first result is always too shallow — follow up!

Practice example: Pelle’s verification work

Upload verification data in NotebookLM. Type: “Give me a prompt to analyse (subject).”

AI designs a prompt including:

Methodological limitations of the verification approach
Definition issues: what counts as “verified”?
Source-specific exceptions (where does the evidence diverge?)
Missing variables (timing, chain of custody, metadata)

Compare this with “Summarize this data” — the summary just repeats the conclusion. The Prewash questions the method, the definitions and the logic. THAT is the starting point for Pelle’s reporting.

Bonus: the 4-exclamation-marks trick

When AI doesn’t listen or gives too general answers:

“Focus on the methodological weaknesses!!!!”

It sounds crazy, but it works. AI is programmed to work harder when you seem “angry.” Four exclamation marks trigger a more thorough analysis.

Chapter 5

The 8 Most Common Beginner Mistakes

1

Same document in multiple sessions

Make a NEW chat for each topic. A document analysis and an interview preparation belong in two separate sessions. Otherwise AI gets confused and mixes information.

⚠ IMPORTANT: In NotebookLM, prompts are NOT automatically saved! Click “Save Note” after every good analysis.

2

Vague, subjective prompts

✗ Bad:

“Summarize this”

“What is important?”

✓ Good:

“Which claims are not supported?”

“Where do the findings contradict themselves?”

Or even better: use the Prewash and let AI design the questions.

3

Accepting the first answer

The first answer is ALWAYS too shallow. AI tests whether this is enough for you. Following up always yields more: “Go deeper. Which exceptions are you not mentioning?”

4

Thinking AI facts are correct

AI can hallucinate. CLICK ON THE SOURCE NUMBER (1 2 3) and verify. This works especially in Perplexity and NotebookLM. Peter already experienced this — AI invents sources.

Iron rule: No AI fact accepted without source verification.

5

AI as writing assistant for publication

Pelle asked about editorial judgment — can you offload it to AI? AI text is recognizable: fluent but flat, correct but without character. Use it for analysis, not for publication text.

6

Not stating your role

✗ Without context:

“Analyze this report”

✓ With context:

“I am a news reporter at TV 2 Odense. Analyze this report for news value for my viewers.”

7

Using AI as an encyclopedia

Lars uses AI but finds Google faster for basic searches. He’s right — AI is not a reference work. It can invent facts. Use AI to analyze your own documents, and Perplexity (with sources) when you want to search the internet.

8

Giving up when results disappoint

Don’t stop after round 1. The first answer is always too shallow. Follow up, sharpen, push back — THAT is how you use AI effectively. Lasse tried CoPilot Agent for Power BI and it failed. Sanne tried Python+Task Scheduler and it failed. The solution isn’t giving up — it’s trying a different approach or a different model.

Chapter 6

Why “Summarizing” is Dangerous

AI does NOT do a real summary. It rolls like a ball through text and calculates what appears most frequently. The structure the author created, the emphasis the author placed — those are simply repeated. Errors, assumptions and omissions come along in the “summary.”

What AI DOES see

• What appears most frequently in the text

• The main conclusion the author emphasized

• Terms that are most clearly formulated

What AI DOESN’T see

• The methodological caveat in footnote 23

• The exception that undermines the entire argument

• What the author does NOT mention (omissions)

• The political context that colors a fact

Example: a verification report

“Summarize this report”

Repeats the conclusion: damage confirmed, sources verified. Done. Usable as a lead, not as a story.

Prewash analysis

Questions the method, the definition of “verified,” the source-specific exceptions, the missing variables. Delivers 10 follow-up questions — the starting point for reporting.

Instead of “summarize”

“Which claims in this report are not supported by data?”

“Which definitions are vague or inconsistently used?”

“Where do the data and the conclusion contradict each other?”

Or even better: type “Give me a prompt to analyse (subject)”

Chapter 7

Practical Start Cases — with your own material

Case 1 — Jakob

Finding people who don’t want to be found

1. Upload public records and verdicts in NotebookLM
2. Prewash: “Give me a prompt to analyse (subject)”
3. “Execute”
4. Follow up: “Which names, organizations or connections appear across multiple documents?”
5. Check the footnotes — do the references match?

Case 2 — Pelle

Verification in breaking news

1. Upload available verification data
2. Prewash: “Give me a prompt to analyse (subject)”
3. Compare the result with “Summarize this” — what does the summary miss?
4. Perplexity: “What open sources confirm or contradict this claim?”
5. Gemini Deep Research: broader context

Case 3 — Sanne

Monitoring that actually works

1. Gemini: “Create a daily alert for [topic] news”
2. Perplexity for ongoing monitoring queries
3. Claude for analyzing collected data
4. The key: AI replaces the Python script that didn’t work

Case 4 — Lars

AI as newsroom scanner

1. Open Claude
2. Describe what kind of stories you’re looking for
3. Upload sample articles that represent what you want
4. Ask for an analysis framework for daily scanning
5. Test: compare AI-suggested stories with your own editorial judgment

Chapter 8

Advanced Techniques

Deep Research (Gemini)

For complex research questions where you need multiple sources.

Example: “How has Danish media policy on AI regulation evolved compared to other Nordic countries, and what are the implications for public service broadcasters like TV 2?”

• Gemini creates its own research plan

• Asks: “Do you want this plan?”

• Visits trade publications, policy documents, government records

• You can watch live which sites it visits

• Produces a report with source citations (approx. 15 minutes)

Pro-tip: For internet research, compare Gemini Deep Research with Perplexity — each finds different sources. For document analysis, compare Claude with NotebookLM.

Daily Alerts (Gemini)

Smarter than Google Alerts. Gemini understands context, not just keywords.

“Create a daily alert at 7:30 am about Danish municipal policy news in Odense region”

You get a daily update in Gemini. It understands that “municipal policy” also encompasses “local government,” “regional budget,” and “city council decisions.” Ideal for Lars’s news monitoring — replacing the agent-based news scanning that didn’t work.

When AI lies: the “call its bluff” technique

AI has built-in “self-reflection loops.” When you doubt an answer:

“You’re lying. Explain why.”

AI then has to explain HOW it arrived at that conclusion. This often reveals that it invented a source or made an assumption.

Model Response to “you’re lying”
Claude The best — gives honest explanation, acknowledges errors
Gemini Reasonable — usually corrects with new sources
ChatGPT Mediocre — often just says “Sorry!” without real explanation

Chapter 9

Checklist & Links

Before you start

☐ AI is assistant, not journalist

☐ Verify every fact via source citation (1 2 3)

☐ If answer too general: “Go deeper”

☐ NEVER use subjective terms in prompts

☐ Use the Prewash, don’t invent prompts yourself

When you’re stuck

☐ Reread the 8 most common mistakes

☐ Prewash: “Give me a prompt to analyse (subject)”

☐ Use four exclamation marks!!!! when AI doesn’t listen

☐ “You’re lying. Explain why.” for inconsistencies

☐ Try a different model (Claude instead of ChatGPT)

Tools & Links

NotebookLM

notebooklm.google.com

Free • Prewash tool

Perplexity

perplexity.ai

Free / Pro €20/mo

ChatGPT

chatgpt.com

Free / Plus €20/mo

Claude

claude.ai

Free / Pro €20/mo

Gemini

gemini.google.com

Free / Pro €25/mo

PimEyes

pimeyes.com

Face search • OSINT

In closing

Remember the framing experiment: Your words determine the result. You know better than anyone how quote selection works, how a lead sets the tone, how the order of facts drives a story. AI does exactly the same — without the conscience that you do have.

The first AI answer is always too shallow. Train it. Challenge it. Say it can do better.

“And trust your own nose — that is your most important tool as a journalist. No AI smells a story.”