How articles are made and checked

The eight rules every page follows

You can check each of these rules yourself on any page. The date next to a claim links to rule 01.

1. 01

   A platform claim carries the date it was checked

   Platform rules change about every three months. So every claim about a platform shows the date we last checked it, right next to the claim. Do not rely on a claim without a date, including ours.

2. 02

   We quote the platform itself, not a summary

   When we quote a rule, we take it from the platform’s own help centre, newsroom or terms, or from the law itself. News sites and blogs only tell us that something happened. If they say something different from the original, we follow the original and say so.

3. 03

   If we could not check it, we say so

   We show gaps as gaps. The label matrix lists the points nobody could confirm, and that list stays on the page. A table that looks complete but is partly guessed is worse, because someone will act on the guess.

4. 04

   Every article names something that did not work

   For example a training run that went wrong, a setting that made things worse, or a workflow we gave up after two weeks. Sometimes the problem was in the research: a number no company publishes, or a rule nobody has written down. Those blocks are labelled “What the sources do not say”, so you can tell them apart from our own experience.

5. 05

   We add corrections, we do not hide them

   When something turns out to be wrong, the original text stays. We add a correction with a date underneath it. That way you can see what was wrong and when we fixed it.

6. 06

   A number that changes is kept in one place

   Revenue shares, waiting times, fees and similar numbers are written in one place, with a date. Other pages link to it. If the same number were in five articles, four of them would be out of date within a year.

7. 07

   A person checks everything before it goes live

   AI may help write a draft. Publishing is done by a person. A person reads, edits and approves every article, and every change is saved in the git history as its own entry. The process is described below.

8. 08

   An automatic check stops broken pages

   An automatic check runs every time the site is built. It checks headings, links, image descriptions, dates on sources and more, and it stops the build if something is missing. It cannot tell whether a sentence is true. It does catch the technical mistakes.

The pipeline, step by step

Status on 30 September 2026: the process below has been running since 22 September 2026. Drafts are made in the operator's own content system. Each draft arrives in this site's repository as a pull request and passes an automatic check (structure, sources, links, the failure block). Then a person reads, edits and merges it. Articles published before that date were researched and written by hand, with AI help in some cases. A person edited those too, but without the pull-request step. We update this paragraph when the process changes.

1. A question comes in. Most questions come from the Discord. A script collects the questions that are asked again and again in the support channels.
2. The content system writes a draft with a language model and a fixed style guide. The style guide bans filler phrases. It requires a date and a source for every claim about a platform, and at least one thing that did not work in every article.
3. The draft is checked automatically. A draft without a topic, a main keyword, a publish date and FAQ entries is rejected before a person opens it. Every new entry starts as draft: true.
4. The draft becomes a pull request on its own branch. Nothing is on the live site yet, and no automation can merge it.
5. A person reviews and edits it. This is the most important step. What the review covers is below.
6. The merge puts it live. A person always does the merge.

What the review covers

The review is more than reading the text once. Every review does the following. A draft that does not pass is rewritten or thrown away:

• Every claim about a platform is checked against the original source: the policy page, the changelog or the terms. The date of the check goes into the article.
• Every number is traced back to our own test or to the source that published it. A number we cannot trace is deleted.
• The failure the article reports really happened. An AI model can invent one that sounds real, and an invented failure is worse than none.
• Legal statements are checked against the text of the law, and they stay marked as not being legal advice.
• Structure, order and wording are edited by hand. Most published articles differ a lot from their first draft.

There is a second reason for this review, besides the law. Search engines watch this niche closely for low-quality pages. A site that publishes AI text without checking it gets ranked down quickly. So the review is good for quality and for compliance.

Who is responsible

Gus (Operator, AI OFM City) reviews and merges every article on this site and is responsible for what is published. More about that role is on the about page.

Where the record lives

In the git history of the repository. Every change to an article is saved there: the text that changed, who changed it, and when. It is a record of what happened, not a description written afterwards.

To be exact about the current state: articles published before 22 September 2026 were written and saved directly by one person on the main branch, and the history shows that. Since then every generated draft has arrived as a pull request, and a person has merged it after the automatic check passed. There is still no second reviewer: the person who edits is the person who merges. The pull-request step does not apply to older articles. We update this paragraph when either of these changes.

Two rules always apply, because breaking them would destroy the proof that a review happened. The history is never squashed. And a generated draft never goes live without its own review commit.

Why we publish this

Article 50(4) of the EU AI Act has applied since 2 August 2026. It says that AI-generated text that informs the public about matters of public interest must be labelled as AI-generated. There is an exemption: the label is not needed if a person reviewed the text and a person or company holds editorial responsibility for it. The European Commission's guidelines of 20 July 2026 read that exemption narrowly. A quick look is not a review, and it must be clear who is responsible.

It is not settled whether articles about platform rules and EU law count as "matters of public interest". We do not want to rely on that question, so we have the review, the named person and the record in any case. This page describes our own process. It is not legal advice and not a template for you.

Images

This site shows realistic AI images of people. The figures on the homepage are generated characters. None of them shows a real person. Article 50(4) still requires a label for them. An earlier version of this page said the opposite, and that was wrong.

So each of them has a visible label wherever it appears. The visible label is what the law asks for. It does not depend on data inside the file. About the files themselves: the images on this site are WebP copies exported from the generator's originals, and the short video clips were cut out and re-encoded for the web. Neither export kept the C2PA manifest. An earlier version of this page said the metadata stayed untouched, and that was wrong. We keep the originals off the site as proof of what was generated and when. If you publish as a persona, keep the AI markers in your files. The AI Act does not ban you from removing them, but platforms read them to label your posts. People in this field lose them by accident all the time, usually through a re-encoding step like this one. This article explains how to keep them.

Image metadata can also reveal who you are. An export from Canva or a generator often contains an author name, an account ID and a document ID. One of our images went live with all three before we noticed. Now an automatic check rejects any image under public/ that contains EXIF, XMP or IPTC data. The only exception is a C2PA manifest, which is allowed to stay. Diagrams, screenshots and abstract graphics are not affected, because they never had such data.

What this does not cover

• We are not lawyers. Everything about EU rules is our reading of the published text. Use it to know which questions to ask a lawyer.
• What we say about how platforms behave comes from a small number of accounts. It may be different for a bigger account or in another country.
• Some pages have affiliate links to our partners, Fanvue and Shift Proxies. Every such page says so at the top, every link is marked "affiliate", and every tool review says who the tool is not right for.
• Nothing here tells you what you will earn.

Corrections and updates

When a platform changes a rule, we update the article and show the new date on the article itself. If you find something wrong, say so in #question-thread: corrections from operators who hit the problem first are the fastest signal we have.

When we get a fact wrong, the article says so in a dated note and in its History list. Every correction on the site is collected on the corrections page.