How this site uses AI

This journal is made by one person working with an AI research system. Here is who does what, what keeps it honest, and why it is built this way.

Who does what

W.F. is a systems designer, with a long habit of finding where systems break, and a long-standing interest in geopolitics. Not a journalist, not an editor, not a specialist in energy or trade.

The AI does most of the legwork. It searches for sources and saves copies of them (never of paywalled pages), logs events and data in the research database, grades how each event bears on each explanation, drafts the articles, translates them into French, and runs the checks before anything is published.

The author directs and decides. The author sets the questions and the odds, reads every draft, argues with it, sends it back for changes, and decides what is published and when. Nothing goes public without that approval.

What the AI is

Not a general-purpose chatbot left to write on its own. It is a research system the author built and also uses professionally, for regulatory compliance and reverse engineering. It keeps a structured record of the project, works from saved copies of its sources, and has to show where each claim comes from. The language models underneath it are commercial models, and they change over time as better ones become available.

What keeps it honest

• Every claim is tied to a source, and the source text is saved so it can be checked again. Paywalled articles are the exception: for those we keep only the headline, date and link.
• Quotes are checked against the source text. Something the author heard in conversation is treated as a lead until a public source is found.
• Records are reviewed before they are made public. Those reviews regularly catch mistakes in the AI's work (a wrong date, a paraphrase shown as a quote, a judgment that did not hold up on a closer reading), and they are fixed before release.
• The grades of how each event bears on each explanation (fits, against or neutral) are made by an AI model working blind. It sees the event, its saved sources and the definitions of the readings, but not earlier grades, our arguments or which reading anyone favours. The author reads them, but they are the machine's judgment, kept that way so that our own leanings don't shape them.
• Events that rest on weak or one-sided sources carry a caution label.
• Once the site is live, every correction goes in the public corrections log, with the old wording and the new.

Our copies of sources

The system keeps a private copy of each page it reads, so it can check a source again quickly without putting more load on the publisher’s servers. Those copies are for the machine only: neither this site nor the private research database shows them to readers, and every citation sends people to the publisher’s own page. Paywalled pages are the exception: we keep no copy of them at all.

Paywalled sources

Many news sites send the full text of a paywalled article to every browser and simply hide it on screen. We respect the protection they chose, including the part done in the browser. The research system reads paywalled pages only in a real browser, as a visitor without a subscription, with the page's own scripts running and its paywall working as intended, and it reads only what that visitor sees on screen, which is often just the headline and byline. It saves none of it: for a paywalled article we record only the headline, the date, the outlet and the link. It never downloads the raw page, never reads the hidden text, and never uses archive copies. That rule is written into every research task, and the system declines when asked to break it.

Every paywalled source is flagged in the database, and we never quote them. Facts come from free or primary sources, and a paywalled outlet is cited as corroboration (it reported the same thing), with a link that sends you to it. When a paywalled outlet is the only source, we state the fact in our own words and credit the outlet, going no further than what a visitor can see, usually the headline. The article editor marks in red any quotation that rests only on a paywalled source, warns the author before an article with such a problem is published, and flags every paywalled citation until the author records that it was reviewed.

A login is a lock too. Some outlets, such as Le Droit, Le Soleil and the other papers of the same cooperative, ask readers to sign in with a free account before showing an article. That is not a paywall, but it is the same kind of protection, so we treat it the same way: we don’t sign in, we read only what a visitor who hasn’t signed in sees, in a real browser, we keep no copy, we quote nothing from those pages, and the facts come from free or primary sources, with the outlet cited as corroboration.

We think the design is the problem. Many paywalled sites send the whole article to every browser, hide it with a script, and then take AI companies to court over what their software collected. Several outlets we cite are doing exactly that: The New York Times is suing OpenAI and Microsoft, and a group of Canadian publishers including The Globe and Mail, the Toronto Star, CBC/Radio-Canada, Postmedia and The Canadian Press is suing OpenAI. Both suits allege, among other things, that the publishers' protections were circumvented. Whatever the courts decide, a page that ships the full text to every visitor is not a lock. It is a curtain, and suing is not the same as fixing it.

So we hold ourselves to the line their design does not enforce, and we ask for the obvious fix: a feed or interface that gives software what a visitor without a subscription can see (headline, byline, date, summary and link) and nothing more. The Globe and the Times do publish such routes, but license them for personal use, not for a project like this one. Yet since 2025 the Times has licensed its journalism to Amazon for AI use, reportedly for US$20 million to US$25 million a year. The route exists; it is sold to the largest buyers. So for them, as for every paywalled outlet, we use only what they make public to anyone: the headline, the date, whatever else a visitor can see, and a link that sends you to them. Every outlet we cite should offer a route software may actually use, and we would use it for all of them.

Our own data, for software too

We offer what we ask of others. Everything public on this site is also published as plain data files that any program may read, with no account and no key: events and how each one bears on each explanation, sources, claims, odds, predictions, projects, laws, actors, data series, and the list of articles. The open data page describes each file and how to read it. They are rebuilt with every update of the site. Our own records are free to reuse under CC BY 4.0, with credit to Carnet 49. The data series belong to their sources, and the articles themselves stay ours. The files carry no quoted text from other publishers: for each source, only the headline, outlet, date and link, as on the site.

The limits

AI systems make mistakes, and they make them fluently. One reviewer without specialist training catches many of them, not all. That is why every claim links to its source: you don't have to take our word, or the machine's, for anything. When an error is found, it is corrected in public.

Why it is built this way

Not by preference. The journal has no funding, no newsroom and no staff. With the means, it would be done differently: journalists and an editor writing and editing, analysts running the research desk and doing their own digging, and the AI as their assistant. Until then, this is one person and one system, doing their best, in the open.