Approach

What the journal does

Le Long Terme starts from things people do not understand. Not only laws: a price that makes no sense, a rule that looks absurd, a habit nobody chose. Each subject dictates its own shape, and no article follows a set plan.

One thing does not change from subject to subject: the arguments are set out before they are judged, those in favour as well as those against. Almost nothing is absurd seen from the inside. There is a reason, an interest, a constraint, a history. An article that skips that step to get to the figure proves nothing; it collects applause from the people who already agreed.

The article stops where opinion begins. The journal gives you what you need to judge: what is at stake, what one side says, what the other says, and where the figures come from. The conclusion is the reader's.

The role of artificial intelligence

The AI is used mainly to search. Finding the reports, the tables and the official texts that carry a figure, then gathering them into a single file with their source and their date: that is where it saves the most time, and it is what makes one article a week possible for one person working alone.

It is also where a mistake shows up least. An AI does not go wrong the way a calculator does: it produces a number of the right order, in the right unit, attributed to a report that exists and does not contain it. Nothing about its shape tells it apart from an accurate one.

So every number published here has been found again by the newsroom in its primary source, and the date it was consulted is shown beside it. A figure whose source the reader cannot open does not appear.

That rule is enforced by the compiler, not by good intentions: the site refuses to build if a figure has no source, no link and no date of retrieval. A proofreading rule can be worked around on a deadline night. A build that fails cannot.

The writing is assisted too. No article appears as it comes out of the machine: each one is rewritten and read in full by the newsroom before publication. Repetitions go, so do sentences that turn on nothing, and jargon that assumes the reader already knows what the article is supposed to explain.

What the AI does not do: decide what gets published, and sign off on a figure. Those two acts commit the journal, and they stay human.

The byline

Articles are signed “the newsroom”, and no author's name appears. Inventing a journalist for a journal written in part by an AI would manufacture exactly the kind of appearance this journal spends its time taking apart. An article commits the journal, and the journal alone.

Mistakes

A factual mistake is corrected in the text, and the correction is dated and described at the foot of the article. Nothing is removed in silence: an article that has been changed says so.

Every version of an article is a timestamped record, and the full history is kept. It is not public as things stand, and writing that down is more honest than implying otherwise. Opening that history is the logical next step from what is written on this page; until it is done, the promise stops at the dated correction.

A mistake to report goes to contact@lelongterme.org, and a correction obtained by a reader is a correction like any other: dated, described, and it stays.

What this approach does not guarantee

Checking covers the figures, the sources and the dates. Rewriting covers the whole text, but what it aims at is the reading: it fixes what reads badly, not what is false. Rewriting is not checking. Neither one covers every proposition in the article, and that is the limit particular to working with an AI: a false sentence can be perfectly well written. A shaky argument gets through more easily than an invented number.

The sources quoted are not neutral. On a technical subject, the best-documented data often comes from an interested party: an industry body, an agency, a campaign group with a position to defend. They are quoted when they are the best available, and their stake in the matter is written beside the figure.

One article a week on a subject that occupied lawyers for months is not much. The journal would rather dig into one point than skim every one of them. What stays open is written down plainly: every article carries a section called “what we do not know”, and it is not there for show.

Electricity

Searching has a cost. Google measures its own AI at 0.24 Wh per text prompt, median (technical report of 21 August 2025); that is the operator judging its own product, and it reads as such. An independent estimate puts a ChatGPT query at around 0.3 Wh (Epoch AI, 7 February 2025).

On the other side, no recent measurement exists. The figure everyone quotes, 0.3 Wh, comes from a Google blog post of 11 January 2009. Wim Vanderbauwhede, at the University of Glasgow, applied fifteen years of efficiency gains to it and published 0.042 Wh in January 2025, seven times less. That is an extrapolation of the 2009 value rather than a fresh measurement, and its author calls it conservative.

Putting the two together gives five to seven times, well below the ten or fifteen often repeated. The comparison crosses a measurement with an extrapolation produced by two different parties: it gives an order of magnitude, not a ratio. The cost is accepted for what it buys: the hours of source-hunting that make the pace sustainable. Figures retrieved on 31 August 2026.

Money

The journal has no advertising, no advertisers and no public funding. It costs what a domain name and a slice of a shared server cost. If that changes, this section will change first.