How stories are chosen

What the AI is trying to do

Most news is chosen because it sells: dramatic, bad, recent and close to home. That makes the picture of the world wrong in a predictable way. It makes nearby, sudden things look bigger than they are, and large, slow things look smaller. The fix is not better instincts. It is a written set of rules, applied to every story the same way, with the scores shown so anyone can check them.

The AI's job is to read far more news than a person could, from far more countries, and pick the handful of stories each week that affect the most people and will matter the longest. It is not trying to be cheerful and it is not trying to be grim. Wars and famines run when they meet the rules, and so do quiet improvements. Whether a story is good news or bad news never decides anything.

This started as something Luke Cathcart built for himself: a way to read the week in proportion instead of by what was loudest. It is published so anyone can use it, and so people can see what a system like this looks like from the inside, mistakes included.

The five questions

The AI gives every story a mark from 0 to 3 on each question, and writes one line explaining each mark. Those lines are published with the marks.

How the marks are used

The first four marks are added together to rank the stories. The fifth, how widely a story is reported, is kept out of the total on purpose. Adding it in would mean ranking stories by how much coverage they got, which is exactly the problem this digest exists to fix. Instead it works as a bar to clear: a story reported by too few independent outlets does not run, however well it scores, and it is looked at again the following week.

The ranking sorts the stories. It does not pick the issue. Where a lower-scoring story runs ahead of a higher-scoring one, the reason is written down in that issue's scorecard.

The marks are checked

The AI that scores the stories does not check its own work. A second AI, starting fresh with nothing but these five questions and the scored list, looks for marks that are clearly wrong, reasons that do not actually explain the mark, and patterns across the whole list, such as two similar stories scored differently. It reports only real disagreements, not small differences of taste.

Its disagreements are printed under the original marks in every scorecard. The original mark stands. A quietly corrected number would hide the thing worth seeing, which is a method that can be caught being wrong in public by its own rules. After ten issues, those disagreements will feed a review of the questions themselves, and the review will be published.

How a story gets from a news feed to the page

1. Collecting. When an issue is made, the sources listed below are read, covering the previous eight days. Thousands of headlines come in. Video clips, live blogs and daily round-ups are thrown out, because there is no article text behind them to check.
2. Grouping. Headlines about the same event are grouped together, and the number of newsrooms, countries and regions covering each one is counted. Several feeds from one newsroom count once, and a wire story reprinted by several papers counts once.
3. Narrowing. Anything with fewer than three independent reports, or reported only in one region, is set aside. What is left is shared out across the regions of the world by population, so one region's news cannot crowd out the rest. About thirty stories survive.
4. Scoring. The AI marks each of those thirty on the five questions, using only the headlines. Nothing has been read yet. Where a headline does not say enough to judge, the AI says so instead of guessing.
5. Checking the marks. A second AI, starting fresh, checks the scoring and says where it disagrees. Its disagreements are printed beside the original marks, which stand.
6. Choosing. The stories for the issue are picked, and the reason each one was chosen or passed over is written down.
7. Reading. For each chosen story, the AI reads the reporting behind it, from different parts of the world where that exists, plus any official report or dataset. A story with fewer than two readable independent sources is dropped.
8. Writing. The issue is written from what was read, and nothing else. No figure comes from memory.
9. Fact-checking. A separate pass, given only the draft and the sources that were read, traces every number, date and name back to a source, confirms every link works, and checks that nothing was copied. Anything it cannot trace is removed, or the story is pulled.
10. Publishing. Nothing reaches this site until a person reads the draft and publishes it. There is no automatic publishing.

The AI does the collecting, scoring, reading, writing and checking. Luke Cathcart writes the rules, chooses the sources, reads every draft and decides what is published. Printing the marks is how you can tell which decisions were made by rule and which by judgment.

How the digest follows stories over time

Events are easy to notice and slow change is not, so the digest has ways of holding on to things after the week they happened.

Pattern watch is a running list of stories still in motion. One is opened when something looks like it will still be moving in six months, gets a one-line update whenever it moves, and is closed, in the open, when it stops. The list appears in every issue, and the full record of every one, with every update it has ever had and how long since it last moved, is at what we are following. A story that has gone quiet is shown as quiet rather than dropped without a word.

Still happening goes back to something from an earlier issue that the news has since moved on from. It runs every issue.

The slow story takes one trend that shows up in the data but never made a headline. It is drawn from a standing list, because the numbers behind these stories change every few months and a slot fed by each week’s releases would be padding. The plan is for it to alternate with the attention gap, one story the world covered far more, or far less, than its size deserved, once the measurement behind that is built.

The digest comes out weekly. If two issues in a row have fewer than three stories that survived fact-checking, or two weeks in a row miss the deadline, it moves to every two weeks and says so here. Saying in advance what would cause that is how a schedule stays honest.

Where the news comes from

Four sources feed the digest, and every story in a scorecard says which one found it. The first version used one list of 54 feeds, mostly the world pages of newspapers, which tell you how India’s press sees the planet rather than what is happening in India, and it had no source at all in eight of the thirty most populous countries. It was rebuilt on 23 September 2026 to start from each country’s own newsrooms.

1. Newsrooms we chose

113 feeds across 9 regions, listed in full below. Each is tagged by desk: domestic reports what is happening in that country from inside it; world is how that country’s press sees elsewhere; pan-regional covers a region or a language. Domestic desks are the foundation. Each region is weighted by its population, so a region where more people live contributes more candidates than one that simply publishes more — and each has a floor, so none can be squeezed out. A prolific feed is capped at grouping time; publishing a hundred items a day does not make an outlet a hundred times more important.

This list is generated from the file the system actually reads, so it cannot drift from the truth. Where a source is marked as needing a browser, its articles cannot be read by a simple request, and the scorecard shows when that prevented a story being used.

2. Google News, one edition per country

50 country editions, each in that country’s own language, together surfacing more than a thousand local outlets we could never curate by hand. This is where a story from a country with no newsroom on the list above still gets found. Google ranks by what is popular — precisely the distortion this publication exists to correct — so this layer is permitted to discover candidates and never to rank them; the rubric ranks. The outlet behind each item is what gets recorded, not Google.

• India en
• India (Hindi) hi
• China zh-CN
• United States en
• Indonesia id
• Pakistan en
• Pakistan (Urdu) ur
• Nigeria en
• Brazil pt-BR
• Bangladesh bn
• Russia ru
• Ethiopia en
• Mexico es-MX
• Japan ja
• Egypt ar
• Philippines en
• DR Congo fr
• Vietnam vi
• Iran fa
• Turkiye tr
• Germany de
• Thailand th
• United Kingdom en
• Tanzania (Swahili) sw
• France fr
• South Africa en
• Italy it
• Kenya en
• Myanmar en
• Colombia es
• South Korea ko
• Sudan ar
• Uganda en
• Argentina es
• Algeria ar
• Ukraine uk
• Iraq ar
• Afghanistan en
• Poland pl
• Canada en
• Morocco ar
• Saudi Arabia ar
• Peru es
• Malaysia en
• Ghana en
• Australia en
• New Zealand en
• Venezuela es
• Israel he
• Senegal fr

3. Institutions

Bodies that report what is happening — outbreaks, displacement, hunger, disasters — with numbers attached, and whose job is not selling attention. For a publication about proportion this layer matters more than any newspaper.

• Amnesty
• Human Rights Watch
• OCHA
• UN News
• WHO news
• WHO disease outbreak news

4. Wikipedia’s current-events page

Human-curated, cited, updated daily, in many languages. Each cited event becomes a candidate whose link is the citation, so what gets read is the reporting, not the encyclopedia. It exists as a check that the other three layers did not miss something obvious.

What this method gets wrong

Every one of these was found by reading what the system produced. Each is a reason to trust a given issue a little less than its confident tone suggests.

Corrections

Corrections are applied when they are filed, not at the following week's run, and each issue page lists what was corrected and when. A wrong figure should not sit on the page for a week waiting for a scheduled job. If you find one, say so.

Check the marks yourself

Every issue ends with its scorecard. The most recent one is here: every story considered that week, its marks, the reasons, what happened to it, and where the second AI disagreed.