Posted on by nura

GenAI experiment by De Tijd

In a bold experiment, Belgian AI company Techwolf teamed up with national newspaper De Tijd to generate an entire weekend edition using generative AI. Not a mockup. Not a prototype. A full, print-ready 60-page edition.

This was not just a showcase of GenAI’s creative prowess. It was a proof of concept for what happens when you combine workflow automation (in this case via tools like n8n) with large language models (LLMs). The experiment provided deep lessons—both inspirational and cautionary—for newsrooms, digital agencies, and AI technologists alike.

The setup: n8n meets GenAI

Techwolf didn’t throw ChatGPT at a PDF template and call it a day. Instead, they constructed a set of structured pipelines that simulated how real editorial teams work:

  • An editorial workflow ideated topics, pitched angles, drafted stories, and fact-checked content.
  • A layout pipeline assembled the articles, images, and headlines into real print-ready pages.

This was made possible through n8n-style orchestration. GenAI alone doesn’t produce coordination; it just produces text. By creating human-designed workflows and assigning different GenAI agents to various steps (headline generation, interviews, imagery, fact-checking), they mimicked the structure and discipline of a real newsroom.

What worked: Speed, cost, and comprehensiveness

The entire process—from idea to PDF—was completed in less than 24 hours. Total cost? Around $100.

  • Articles were written and fact-checked in 1–2 hours.
  • Layout generation took roughly 4 minutes per page.
  • Images were created using Google Imagen.

Even seasoned editors at De Tijd admitted: “At a glance, it looks like a normal paper.”

Editorial Lessons for Newsrooms

For top newspapers, the experiment raised questions that cannot be ignored:

  1. Speed vs. accuracy: AI can cut production times dramatically, but who guarantees correctness?
  2. Standardisation: AI loves templates—but real stories demand unpredictability.
  3. Relevance: Prompts can prioritize what the model “thinks” is interesting, not what the public truly needs.
  4. Bias in data sources: If models rely on English-speaking sources or outdated facts, how do you localise or contextualise them?

What stood out most was what was missing. For example, all ‘journalists’ were named “Lars” and the layout felt formulaic. That can be fixed. More importantly, the interviews were stilted and lacked confrontation, and the articles avoided complexity or editorial voice.

Trust, subtlety, and critical thinking are what make journalism journalism. And in every example, the GenAI version missed the mark.

n8n + GenAI = Best of both worlds

This experiment showed the value of combining automation (via tools like n8n) with GenAI. n8n workflows bring order, traceability, and logic. GenAI adds creativity, summarization, and natural language generation.

Together, they can massively scale content production.

But this is where we must tread carefully.

Four key risks that must be managed

  1. Prompt manipulation: If a prompt is altered, accidentally or maliciously, the AI may fabricate, distort, or bias content (see our blog post)
  2. Loss of editorial integrity: Without human editing, the final output risks sounding plausible but being wrong.
  3. Hallucination of facts: Even with fact-checking loops, LLMs can generate “almost true” claims.
  4. Oversimplification: AI tends to strip nuance in favor of clarity. Great for FAQs. Dangerous for investigative journalism.

So what now? Practical takeaways

  1. Redactions should experiment safely: Run internal GenAI experiments using sandboxed workflows to learn what works.
  2. Keep humans in the loop: Make final human editorial review mandatory before publication.
  3. Adopt transparent pipelines: Like n8n, use tools that document what happened, when, and why.
  4. Design prompts responsibly: Prompts are like editorial briefs—guard them accordingly.

Final thoughts

Techwolf and De Tijd showed what’s possible. But they also showed what’s missing.

This is not a debate between humans and machines. It’s a new discipline: editorial engineering. Where workflows, language models, and journalists work together—each respecting the other’s strengths and weaknesses.

The newspaper of the future will be built by both human editors and AI pipelines. But the final decision—what’s fit to print—must remain human.

That is not a constraint. It is a responsibility.

References

Leave a comment