Skip to content

AoT Editorial Pipeline

What is Art of Truth?

Art of Truth is an experimental editorial site where the byline is honest about what produced the article: an AI editorial pipeline, supervised by a human.

Every article you read here is researched, drafted, fact-checked, translated, and published by an automated pipeline.

Who builds it?

I’m Maël Guimet. I do RLHF Factuality (Reinforcement Learning from Human Feedback, focused on factual accuracy) for a living for a big company that you know, and I built this site to explore what a responsible, fully-transparent AI-assisted newsroom could look like in practice, not as a thought experiment, but as a working publication with real readers. I face the AI-farm slop issues on a daily basis; this is an honest attempt at building anti-slop.

How the pipeline works

Article generation is a multi-stage process. The short version:

  1. Topic selection. Either trending stories pulled from Google Trends RSS, or topics the editor (me) queues manually.
  2. Research. The pipeline reads primary sources directly, press releases, official reports, court documents, peer-reviewed papers when relevant, instead of paraphrasing other outlets. Sources such as Wikipedia are soft banned, I do not want to compound errors from other aggregators and instead try to rely on primary evidence as much as possible.
  3. Drafting. A persona-specific writer agent (News, Technology, Opinion, True Crime, History, Culture) produces a first draft from the research notes. The persona controls tone, not content. The model behind it is still the same: Claude Opus 4.5, which is the best Opus version for writing style as I’m writing this.
  4. Fact-checking. Each claim is verified against the sources before publication. Numbers, dates, names, quotes, and unit conversions are checked deterministically. This step is critical, and this is where LLMs struggle the most. While I built many mitigation systems and articles are mostly trustworthy, I still want this to be very clear: never trust blindly what an AI says. If something is important, read the cited sources directly.
  5. Translation. Published articles are translated into French, German, and Spanish by the same pipeline.
  6. QA. A QA pass catches unit drift, broken idioms, dropped citations, and other small things that might have gone wrong somewhere in the production line.
  7. Publication. Most articles publish automatically once all steps pass. Articles that hit certain issues are routed to me for manual approval.
  8. External Factchecking. All published articles are manually piped through at least one additional LLM for further verification and QA. Errors are fed back to a dedicated agent, which fixes them.

By that point, we’ve reached a pretty solid confidence level for factuality. But, as I previously mentioned, do not rely on AI for important things. AoT can be trusted to answer your curiosity and give you insights on the topics you care about, but always check primary sources directly for anything actually important. This is also what you should do with something like Wikipedia, although no one actually does it.

What the pipeline is not

  • Not a content farm. Articles are not optimised for word-count or keyword density. They’re generated when there’s something to say. Some SEO junk is still necessary to not get buried, unfortunately.
  • Not anonymous. If you want to reach a human about something published here, use contact@artoftruth.org.

Corrections

If you spot a factual error, please email contact@artoftruth.org.