Skip to content
Our method

Our method: AI writes, fixed rules check

An AI writes fast, but it can be confidently wrong: inventing a figure, attributing to your company a quality it does not have, citing a source that does not exist. Asking it to be careful is not enough. At FaireDuBruit, it is not the AI that writes which decides what gets published: it is rules written in advance, applied by code, and a second AI tasked with reviewing.

The topicThe factsThe writingThe reviewrefusal, with its reasonthe article is published

Three roles, never mixed up

AI writes

It chooses the words, builds the article, suggests an angle.

Rules check

Written in advance, applied by code or by a second AI tasked with reviewing, they accept, have the article rewritten or refuse, and always say why.

Humans set the rules

A rule changes on a human decision, dated and recorded. Our rule: an AI can suggest a change, only a human decides it.

One article, four steps, a check at each one

Every distribution follows the same path. At each step, a check verifies the AI’s work before moving on to the next. If it refuses, the step starts over or stops, with its reason.

  1. 1

    The topic

    The topic and the three media where the article will be published are chosen based on your file and your answer. An unsigned request, or a document that belongs to another file, stops the distribution before any call to an AI.

  2. 2

    The facts

    Before any writing, the code draws up the list of facts the article will be able to rely on: what you have declared, what your website shows (we keep the page address and the passage noted), and your own sentences, used exactly as you wrote them. A piece of information suggested by an AI is not stated on its word alone.

  3. 3

    The writing

    Each paragraph, like the title and the standfirst, must point to the fact or source that supports it; otherwise the article is rewritten, then refused. The AI writes no addresses: sources are placed by code, based on the searches actually carried out. The words you asked us to avoid are refused. The article on your website names none of the competitors we know of (yours and those found by our searches), and the second AI checks that it does not mention any others.

  4. 4

    The review

    A second AI reviews each block of the article (title, standfirst, paragraphs) by comparing it with the facts and the sources. A paragraph that adds something has the article rewritten; if it still adds something, it is removed; if it is the title or the standfirst, the article is refused. If the review cannot be carried out, the article is refused. The English version is reviewed in the same way; the other languages go through automatic checks, without review by an AI. An article too close to another article in the same distribution is rewritten once, differently.

On the last attempt, a paragraph that the review still disputes is removed rather than rewritten once more: we prefer a shorter article to an invented sentence. If the flaw affects the title, the standfirst or a substantive rule (support, word to avoid, competitor), the article is not published.

No member of our team reviews each article before it is published: these rules, written and tested in advance, are what decide. We prefer to say so. And a single remark from you is enough to have the article on your website redone, up to three times.

What the AI that writes cannot do

Change the rules

The AI that writes has no access to them; a rule only changes on a human decision, dated and recorded.

Go live through a shortcut

Our code goes live through a designated channel, which checks before accepting and refuses while stating its reason; a code file changed through any other channel is detected.

Handle the access keys to the models

They are entered by a human, with masked input, then encrypted. Our rule: no key goes through a conversation with an AI; a key that went through one by mistake is revoked.

Choose its provider

For writing, we only use models from three developers: Anthropic, OpenAI and Google. A setting that named another one is flagged.

Checks that are themselves checked

A check must first know how to refuse

Our rule: before trusting a check, we make it fail on purpose, on test cases, and we verify that it refuses for the right reason.

The code in service is monitored continuously

It is compared with the last authorized release; a code file modified outside the designated channel is detected, and the server concerned can be isolated automatically.

Our checks are themselves monitored

A check that stops running is reported as an alert: our rule is never to take silence for good news.

An alert does not fix itself

Checks observe; the decision to correct remains human.

A record for every article

For every article, we keep the initial brief, signed, with the facts selected and the origin of each one; the sources cited; the models that wrote and reviewed; what the review had removed, and the reason for each rewrite or refusal. Every release of our code is recorded in a signed log, which can only be added to.

AI and the law: what we do

The European AI regulation (article 50, applicable since August 2, 2026) provides, in certain cases, for disclosing that content was produced with AI. Here is, in concrete terms, what we do today.

The article on your website says so

At the top, it carries the notice “Article written with the help of AI.”, including under white label.

Our media say so

The footer of each of our sixteen media states, in its language: “Content written with the help of AI, under the editorial responsibility of FaireDuBruit.”

Advertising is presented as such

On our media, our articles are sponsored publications, and the links to your site carry the sponsored attribute.

What we do not ask of AI

It writes only from the facts described above, with no invented reviews or customer references; the review removes anything it adds, and a figure in the article on your website that is absent from its facts has it rewritten.

Text cited: Regulation (EU) 2024/1689 on artificial intelligence, article 50

Trust can be verified

We do not ask you to believe that AI is never wrong. We show you what we do so that it does not publish a mistake in your name.

Our ethicsOur offer

Method as described on November 10, 2026. This page explains our choices; it does not replace legal advice.