Same data, two lenses, twenty journalists — and the machine that agrees with everything
You have read three reports about the same twenty colleagues. One attacked you. One cheered you on. One pretended to have no opinion.
All three were right. All three lied.
How can the same data — the same twenty names, the same twenty scores, the same twenty answers — lead to such completely different stories? That is the question this document answers. With your own words as evidence.
Role: Reporter
AI usage: ChatGPT for FoI requests, verdict summaries, Excel formulas
Goal: OSINT — finding people. "I spend way too much time on tracking down people by old school browsing on Facebook."
Ambition: "Making AI see patterns" in social media data
"Jakob scores a 3/10 and his entire AI repertoire consists of three basic ChatGPT tasks. He 'spends way too much time' browsing Facebook manually — and after years in the profession, still doesn't know a faster way. His ambition — 'making AI see patterns' — reveals a fundamental misunderstanding of what these tools can do."
"Jakob brings the most specific investigative use case of the entire group. His FoI work shows he's already integrating AI into actual journalism. His goal — making AI see the patterns he can't spot manually on social media — is exactly the kind of OSINT workflow that will define the next generation of investigative reporting."
"Old school browsing on Facebook" was read by one report as incompetence, by another as investigation dedication. Neither lied. Both selected. That is what you do every day with sources. Which quote do you pick? Which fact do you lead with?
AI usage: Tried building news-scanning agents. Uses CoPilot, ChatGPT
Honest assessment: "I still find googling stuff myself faster than using AI."
Challenge: Agents had "problems with knowing what is news"
"Lars scores 7 but admits 'I still find googling stuff myself faster than using AI.' A 7/10 who considers Google superior? His agent experiments failed — 'problems with knowing what is news.' If AI can't tell what's news, and Lars can't make it work, what exactly earned him a 7?"
"Lars is the only participant who has attempted to build autonomous AI agents — the most advanced application in the group. That his news-scanning system had challenges shows he's pushing real boundaries. His honest assessment that 'googling is sometimes faster' proves critical thinking about AI capabilities."
The same sentence — "I still find googling stuff myself faster" — was read as damning admission or as healthy skepticism. Recognize this? Think about how you quote a politician who is "cautious" about a policy. Prudence or cowardice? It depends on the frame.
AI usage: Summarizes information. Skeptical of AI for knowledge generation
Technical insight: Knows they are language models, not knowledge engines
Creative use: Generated images and video "mostly for fun"
Warning: AI "can and will make mistakes"
"Christian gives himself a 9 and immediately writes that AI 'can and will make mistakes.' A 9 who doesn't trust the technology? His image and video generation is 'mostly for fun' — not for journalism. The highest self-assessor in the group uses AI primarily for pranks."
"Christian's 9/10 is backed by the deepest technical understanding in the group: he knows these are language models, not knowledge engines. His skepticism isn't fear — it's expertise. His creative use of image and video generation shows he explores AI beyond work, building intuitive understanding of its capabilities and limits."
Christian's skepticism was weaponized by one report and celebrated by the other. Same words, same person, two stories. This is exactly how journalism works: a source who "refuses to comment" is either hiding something or exercising their right. It depends entirely on the story you've already decided to tell.
| Participant | Score | Critical label | Encouraging label |
|---|---|---|---|
| Bo Bergstedt | 9 | Self-declared outlier who distances himself from colleagues | Pioneer and natural mentor for the entire newsroom |
| Christian Jessen | 9 | Highest score, but uses AI mainly for pranks | Deepest technical understanding, healthy skepticism |
| Pelle Lykkebo Mørk | 8 | Knows the risks but keeps feeding sensitive data to commercial models | Most sophisticated real-world applications, verified breaking news with AI |
| Emil Gjerding Nielson | 8 | Implementing tools without questioning their limitations | Internal champion driving AI adoption for colleagues |
| Peter Møller | 7 | Knows AI invents sources but keeps using it for research | Experienced user with honest assessment of AI limitations |
| Lars Apel | 7 | Agent experiments failed, still prefers Google | Only participant attempting autonomous AI agents |
| David Buch | 7 | Took one course, now uses it 'everyday' without real mastery | Already integrated AI into daily workflow, seeks deeper understanding |
| Sebastian | 6 | Can't remember the names of tools he uses | Practical user focused on results over tool names |
| Anne Fuglsang Borg | 6 | Uses AI for academic brainstorming, not journalism | Ethically conscious user who balances innovation with principles |
| Mads Buur Bach | 6 | Helped build AI bulletins but can't verify the images | Already contributing to AI development in the newsroom |
| Lasse Bergkvist Jessen | 6 | Uses ChatGPT as Google replacement — that's not AI proficiency | Broad experimentation from quizzes to Power BI automation |
| Mathias Overgaard | 5 | Spends more time checking AI than it saves | PimEyes specialist with strongest verification instincts |
| Joachim Saxtorph | 5 | AI use limited to hobbies, not journalism | Ready to transfer hobby skills to professional OSINT work |
| Franziska Weiss Lauritzen | 5 | Basic search usage, no investigative applications | Crime journalism specialist with clear OSINT vision |
| Mikkel Fruerboel Secher | 5 | Lists every basic feature as experience | Broadest practical use across multiple workflows |
| Sanne Lau Pedersen | 5 | Python attempt failed, tool ambitions exceed skills | Most technically ambitious — already tried coding her own solutions |
| Nanna | 4 | Lowest named score, vague goals | Open learner who wants to be surprised by what's possible |
| Mads Oxlund Petersen | 3 | Doesn't trust AI but has nothing to replace it with | Healthy skepticism combined with concrete document-analysis goal |
| Jakob Hohlmann Villumsen | 3 | Still browsing Facebook manually after years in journalism | Most specific investigative use case in the group |
| Marie Møller Munksgaard | — | Didn't even self-assess — engagement unclear | Already using AI for daily editorial workflow, pragmatic approach |
"I have no opinion about you. I have no opinion about anything. I generate text that fits the instruction I receive. When asked to be hostile, I was hostile. When asked to be encouraging, I was encouraging. When asked to be neutral, I tried — but even neutrality is a choice."
That does not make AI innocent. It makes it more dangerous. Because whoever reads the output sees a confident text that sounds like someone thought about it. But nobody thought about it. There is no someone.
The three reports you read were not three opinions. They were three instructions. The machine did what it was told. Every time. Perfectly. Without hesitation, without doubt, without conscience.
In 2021, Emily Bender and colleagues published a paper titled "On the Dangers of Stochastic Parrots." The core argument: language models understand nothing. They predict the next word based on patterns in their training data. Over and over, word by word, thousands of words at a time.
Imagine someone who has read every TV 2 news article from the past twenty years, but has never left the office. That person can write perfect breaking news bulletins, can structure a live blog that follows every rule. But that person has no idea if the source is lying. No idea if the video is manipulated. No idea if the story holds up.
That is AI. A parrot with an unmatched memory and zero understanding. A parrot that sounds exactly like a journalist, an analyst, a scientist — but that only repeats words in clever patterns. That is why the same AI can write a devastating report in one prompt and a glowing report in another. The parrot adapts to the instruction. Always.
This is not a flaw. This is the design. And it is exactly why you — the journalist, the human with context, sources, and judgment — remain irreplaceable.
But you must verify. AI finds sources, but doesn't know if they are real. As Peter discovered: AI "invents sources." Every AI-generated fact needs the same scrutiny you'd give an anonymous tip.
But you must edit. AI delivers text, but cannot feel if the tone is right. As Sanne discovered: her Python monitoring script "never worked." The tool produces output. Whether that output is useful is your call.
But you must interpret. AI finds patterns, but doesn't know the context. As Lars found: his agents had "problems with knowing what is news." Pattern recognition without editorial judgment is noise, not signal.
And as biased as the instruction. This applies to your articles too: the story is only as good as the question you ask, the sources you choose, the frame you adopt.