How we write
Every story here is written by AI from what other outlets reported. This page says exactly how, what we check, and what we can’t check yet.
Where the facts come from
We read 49 outlets, from CNN on the left to Townhall on the right. Each one is placed by how its news coverage is usually rated on public media-bias charts. These placements are a starting point, and we’ll publish any change we make to them.
Outlets shown in grey block automated readers, so we don’t read them. We don’t get around those blocks. Their reporting still reaches us when other outlets cite it, and we credit it when they do.
We read politely: one request to a site every two seconds at most, and we follow each site’s rules for automated readers.
How a story gets made
- CollectEvery five minutes we check each outlet’s list of new articles and save the headline, link, time and text.
- GroupArticles about the same event are grouped into one story. We only write a story once outlets on the left, center and right have all covered it.
- List the factsA model reads the articles and writes a fact sheet: who, what, when, where, why and how, every number, every exact quote, and where the outlets disagree. Every line names the article it came from.
- Lock the quotesCode checks each quote against the article it came from. A quote must appear there word for word and be credited to the same person. Then it’s locked, and no later step can change a word of it.
- Write three versionsLeft, center and right versions are written from the same fact sheet. The side decides what comes first, what gets stressed and whose voices lead. It never changes a fact.
- Make it easy to readEach version is rewritten to an eighth-grade reading level, with short sentences and short paragraphs, without changing what it says.
- Publish with sourcesEvery sentence goes up with numbered sources, and every number links to the outlet’s own article.
What code checks before anything goes up
After every step, code checks the work. If a check fails, the model is told exactly what’s wrong and gets one more try. If it fails again, that version is thrown out.
- QuotesEvery quote appears word for word in a source, credited to the same person.
- Numbers and datesEvery number, weekday and month appears in a source. A draft that called October 9 a Tuesday was sent back; it was a Friday.
- SourcesEvery paragraph has at least one source number, and every number points to a real article in the list.
- Our own wordsSentences can’t copy an outlet’s wording. We say it our way and link to theirs.
- The basicsWho, what, when and where are in the first two paragraphs.
- Reading levelGrade 8 or below, an average sentence of 18 words or fewer, and no sentence over 28.
From our first test run, October 11, 2026.
What the checks don’t catch yet
Our checks confirm that a sentence has a source. They don’t yet confirm that the source actually says what the sentence says. In one test, a sentence about what Ukraine’s president believes was credited to outlets that probably didn’t say it.
We’re adding a check that reads each sentence next to the articles it cites and asks one question: does this source back up this sentence? Until then, every source number links to the original, so you can check it yourself. If one doesn’t hold up, tell us.
Three versions, one set of facts
Outlets on the left and right often report the same event with different facts first. We write all three so you can read the one you’d normally skip, and see that the facts underneath don’t move.
- What changesThe order, the headline, what’s stressed and whose voices are quoted first.
- What doesn’tThe facts, the numbers, the quotes and the sources.
Use the Left, Center and Right switch on any story to watch the words change. Sentences all three share stay put, and only the words that differ move.
When we get it wrong
- A factual error gets a numbered correction, like C-014, on the story and on the corrections page.
- New information after publishing gets a numbered update, like U-031.
- We keep the old wording and show it next to the new.
- We answer every error report.
The AI we use
Stories are written by Gemma 3, an open model from Google, running on our own computers. No story is sent to an outside AI service, and nothing is published without passing the checks above.
[To decide before launch: whether a person reads each story before it goes up, and how that’s shown on the story.]