How these articles are written
In one sentence
These articles are drafted with the help of a large language model, then reviewed, edited or rejected by me before publication. I hold full editorial responsibility for them — mistakes included.
The pipeline, concretely
- The topic is mine. A hand-maintained editorial calendar: one French article on Tuesdays, one English article on Thursdays.
- Drafting is assisted. An automation running on my own homelab calls a Claude model (Anthropic) through a local proxy. The prompt is anchored on my competency matrix and on a style guide derived from my earlier writing — not on a generic persona.
- The model is not allowed to invent numbers. The prompt forbids fabricating hardware, benchmarks or costs: it may only use what I give it as fact. That constraint was added after catching the model filling tables with plausible but false data.
- Nothing ships without my approval. The draft lands in a holding folder and pings me. I read it, fix it, rewrite it or bin it. Publication only happens after an explicit approval from me.
- After publishing, the article is rendered on this site from the same source file, and a comment linking to the long version is added under the LinkedIn post.
Why there is no percentage
"60% AI / 40% human" reads well and means nothing: a percentage of what? Words, ideas, time spent, structure? None of it is verifiable, and an unverifiable transparency notice isn't transparency — it's marketing. Article 50(4) of the AI Act requires disclosing that a text was AI-generated or manipulated, never quantifying how much. I stick to what is true and checkable: there is AI in here, there is a human deciding, and that human has a name.
The legal basis
Since 2 August 2026, the EU AI Act (art. 50(4)) requires deployers to disclose that text published to inform the public on matters of public interest was artificially generated or manipulated. The obligation lifts where the content underwent genuine human review and a person holds editorial responsibility — which is my case. I label anyway: I advise organisations on AI governance, and hiding behind an exemption instead of applying the rule I recommend would be a poor look.
What it does not claim
That the text is correct. The notice covers how the text was produced, not whether it is accurate. When I cite a source, it is linked; when I quote a number, it comes from my own infrastructure or a public reference. If you find an error, say so in the comments under the LinkedIn post — it will be corrected and flagged.