Behind the issues2 min read
How five stories are picked from a week of AI news
AI models gather, score and draft the news, and they follow one rule: what can a working person do with this? How it works, and the same test for your own reading.
Every week brings a pile of AI news. Most of it does not change how you work on Monday morning. The Monday issue tries to find the part that does. This post explains how that happens, honestly: most of the work is done by AI models, following rules written for people like you.
Where the news comes from
Every day a script collects new items from a long list of places: news sites and company blogs, Google News searches, a few Reddit communities, Hacker News, and newsletters that arrive by email. That is far more than anyone could read.
The one question
An AI model scores every item from 0 to 1 by one question: how useful is this to a non-technical person who uses AI at work?
Items score high when they are a change you can use or should know about. For example: a new feature in ChatGPT, Claude, Gemini or Copilot, a price or plan change, a connector or automation you can set up without code, or a privacy rule that changes what you may paste into AI.
Items score low when they matter mostly to developers or investors: developer tools, test results between models, funding rounds and rumours.
From item to story
The best-scored items go to a second AI step that acts as an editor. It keeps one story per event, so five articles about the same launch become one story.
Each story is then written from the original articles, which are read in full, not from a headline. Every story ends with a "For you" line: one thing you can do or check this week. A person can choose which drafts go on the News page. When nobody chooses on a given day, the best-scored draft is published.
From stories to the Monday issue
On Friday the issue for the next Monday is built from the stories published in the seven days before it, best score first. When there are more than five, an AI editor step picks five with some spread: no more than two about one company, and a mix of kinds of news.
The draft issue then goes to the person who runs the newsletter. They can swap a story, ask for a section again, or hold the issue. When nobody steps in, the picks go out on Monday.
What this does not catch
- The scores are a model's judgement. A useful story can score too low, and a dull one too high.
- Sources can be wrong. Stories are written from published articles. If the article is wrong, the story can be too. Check anything important with the company's own announcement.
Use the same test yourself
You do not need a newsroom to filter AI news. Next time a headline catches your eye, ask three things:
- What can I do differently after reading this?
- Where is the original source?
- Can I try it today, or is it only a promise?
If you cannot answer, close the tab. You lose nothing.
Read the latest on the News page and see past editions in the issue archive.