The short version

Every article on this site is built with AI assistance. Every article is then
reviewed, fact-checked and approved by me before it goes live. I hold editorial
responsibility for every claim, every number and every chart you find here.

If a number on this site is wrong, that is my mistake, not a machine’s.

What the AI actually does

Each post goes through two phases of specialist agents and then through me. The
first phase drafts, the second attacks the draft, and I decide.

Phase 1, build. Three agents work in parallel. One writes the article. One
writes the R or Python that produces the data and the charts. One builds the
interactive dashboard, where a post has one.

The order matters: the code runs first and emits a frozen dataset. The writer
and the dashboard builder then quote that dataset. Neither is allowed to invent
a number or research its own. This is why the figures in the prose, the figures
on the charts and the figures in the dashboard agree with each other.

Phase 2, review. Three more agents attack the output. One reviews the
writing against a published house style guide. One fact-checks every claim,
formula and citation against primary sources. One QA-tests the dashboard code,
its math and its accessibility.

Phase 3, me. I read the result, judge whether it is worth your time, correct
what is wrong, and decide whether it publishes. Nothing goes live that I have
not read.

What you can check yourself

I would rather you did not take the above on trust.

• The code is published. Most posts carry a collapsible section with the
  full R or Python that generated its charts and numbers. Run it. If my figures
  do not reproduce, tell me.
• Citations point at primary sources. Not at a blog summarising a press
  release about a report. If I cite a study, the link goes to the study.
• Simulations are labelled as simulations. Where a post uses a teaching
  model rather than field data, it says so, in the post, next to the numbers.
• When I cannot verify something, I say so rather than rounding a
  half-remembered figure into a confident one.

Where the AI does not decide

The AI does not choose what I write about. It does not choose what I believe.
The opinions on this site are mine, including the ones I later turn out to be
wrong about. When a post takes a position, that position was argued for by a
person who has spent fifteen years in purchasing and operations and has to live
with having said it out loud.

On the EU AI Act

The AI Act’s transparency rules for AI-generated content (Article 50 of
Regulation (EU) 2024/1689) have applied since 2 August 2026. The rule most
relevant to a site like this one covers AI-generated text published to inform
the public on matters of public interest.

That rule contains an exemption: it does not apply where the content has
undergone human review or editorial control and where a person holds editorial
responsibility for publication. The pipeline described above is exactly that,
and I am that person. So the footer you see at the end of each article is not a
legal confession. It is there because you deserve to know how the thing you are
reading was made, and because I would want to know if I were you.

The relevant texts, if you want to read them rather than take my summary:
Article 50 and the
European Commission’s guidance on the transparency
obligations.

Questions

If something on this site looks wrong, or if you want to know how a specific
post was produced, write to me. I will tell you, and if you have found a genuine
error I will correct the post and say what changed.