AI transparency

I use artificial intelligence to produce part of the content of this site. This page says where, how and how far, and explains how to read the notice shown at the end of every article.

How I work with it

I am not hiding it: I use AI every day, at work and here. A good operations engineer is lazy in the right place, and that is a professional quality before it is a flaw: whatever is repetitive ends up automated, otherwise it ends up rushed. AI is simply the latest tool to join that list, after the shell script, Ansible and Terraform. Treating it differently would be inconsistent.

In practice it drafts a first version, keeps both languages of the site alive, proofreads, and writes the repetitive code nobody takes pride in. What I get back is time on what actually has value: architecture, technical trade-offs, running things in production and testing them. That is where I want to spend my days, and it is exactly what I do not delegate.

The counterpart is simple: I read everything, I correct a lot, and I sign it. An article published here carries my name, not a model's. When something is wrong, that is on me.

The levels

Every article declares, for each of its parts, how far an AI was involved. One scale covers text, code, images and diagrams, from fully human work to content produced entirely by an AI. The value in brackets is the one written in the source file of the article.

• Manual (manual): no AI involved.
• Assisted (assisted): produced by a human, with an AI for proofreading and suggestions.
• Partial (partial): first version produced by an AI, then corrected in detail by a human.
• Generated (generated): produced by an AI, reviewed by a human.

No level goes beyond generated: nothing is published here without a human going through it.

One extra note on code

The level says nothing about whether the commands actually ran, and that is not the same promise to the reader. Code therefore carries a second mention:

• Tested (tested): Run and tested by a human before publication.
• Indicative (indicative): Given as an illustration, not run as such.

What AI does not do

• Pick an architecture or decide between two technologies: that is what this site is about, delegating it would be absurd.
• Confirm that a stack works. Anything presented as tested was tested on my own machines, not in a model's answer.
• Invent a field report. The durations, the outages and the numbers come from what I actually ran.
• Publish. Nothing goes online without a human review.

Where the notice appears

At the end of the text of every article, before the share links. It is declared in the source file of the article, so it lives in the same commit as the text it describes and is updated along with it.

Legal framework

The applicable text is article 50 of regulation (EU) 2024/1689, the AI Act, whose transparency obligations apply since 2 August 2026. It separates two things: marking synthetic content in a machine-readable format, which falls on the providers of the AI systems, and informing the public, which falls on whoever publishes, in a few specific situations: interacting with a conversational system, deep fakes, and text published to inform the public on matters of public interest.

That last obligation does not apply when the content went through human editorial control and someone holds editorial responsibility for it, which is the case for every article published here. The notice is therefore voluntary, and that is precisely why it interests me: transparency is only worth something if it does not stop at the strict legal minimum.

On the French side, law no. 2024-449 of 21 May 2024 extended article 226-8 of the criminal code to content generated by algorithmic processing that represents the image or the voice of a person without their consent, when its artificial nature is neither obvious nor expressly mentioned. No illustration on this site depicts a real person, and generated images are flagged as such.

This page describes a practice, it is not legal advice.

A question

If a notice looks inaccurate or incomplete, write to contact@toulliou.org and I will correct it. How your data is handled is described on the Privacy page.

Last updated: August 2026.