How we work
An AI pipeline writes the stories on this site. This page describes each step, what it checks, what we do not do, and where the system can still be wrong.
In short
We have no reporters and interview no one. We follow technology companies' own announcements and the technology press, gather the reports that tell the same event, and write one short summary of them. Every fact in a summary rests on one of the sources linked at the bottom of its page, and is compared with them before publishing.
How a story gets published
- Collection. Every hour we read the feeds and article pages of the publications listed below. We do not open pages a publisher's robots.txt asks automated readers to leave alone.
- Selection. A language model reads each article's headline and opening and sets aside what is not technology news. Podcast episodes, newsletters and weekly round-ups are left out, whatever they cover.
- Grouping. Reports of the same event are gathered into one story. No language model is involved: the similarity of the texts is measured on our own machine.
- Summary. A model reads the story's reports and writes a summary in its own words. For every fact it must point to a sentence in a source, word for word. Those sentences are only for checking: they are never published.
- Checking by code. A program confirms that every quoted sentence really is in the source it was attributed to, and that the companies, products and figures named are in the sources. A summary that copies twelve or more words in a row from a source is rejected.
- Importance. Each story is scored from 0 to 100 on reach, novelty, the reliability of its sources, how widely it interests readers and how much of its value lies in reading it today. Stories under 60 are not published.
- Second reading. A second model compares every summary that clears that bar with the sources, sentence by sentence: the counts, the dates, the names and the verb that says what happened. If it finds a claim the sources do not carry, or one stated more strongly than they state it, the summary is rewritten; if the rewrite fails too, the story is not published.
- Writing the page. The headline and the why-it-matters paragraph are written from the verified summary and the facts the second reading confirmed. The model at this step never sees a source text. The Turkish page is not translated from the English one: it is written separately from the same verified facts. A last program looks for any figure not in the verified text, promotional words and copied passages.
Where people come in
No editor approves each story. A page goes live on its own once it has passed every check above.
We set the rules, the thresholds and the list of sources, receive a report of the day's stories every morning, and correct or withdraw a page we find wrong. When we change a step we measure it first, including by comparing the scores people give stories with the scores the model gives them.
Opinion and fact
The summary at the top of a page reports what the sources say. The “Why it matters” section is our assessment of it, and it too may rest only on facts that were verified.
Models, pictures and voices
The text is written and checked by Anthropic's Claude models. Anthropic is also a company we cover; its news goes through the same steps as everyone else's.
Illustrations are generated by AI and labelled as such wherever they appear. They show symbolic objects, never the event itself, real people or logos.
Stories are read aloud by an AI voice, or by your device's own voice where we have no recording. The button says which.
What we do not do
- We do not reprint source sentences. You will find no quotes from the articles on our pages; follow the links to read them.
- No story carries a byline, because no single person writes it.
- We do not publish anything that has not been checked against its sources.
Limits
The checks catch most mistakes, not all of them. The second reading is itself a language model and can be wrong, and if a source is wrong, so are we. If you see a mistake, write to [email protected]. Corrections explains what happens next.
Our sources
Every publication cited at least once in a published story, with the number of stories it appears in.
- TechCrunch
- Tom's Hardware
- BleepingComputer
- The Verge
- Ars Technica
- Engadget
- InfoQ
- BBC Technology
- The Hacker News
- Google Blog
- SecurityWeek
- PC Gamer
- Wired
- Kotaku
- MacRumors
- Anthropic News
- GitHub Blog
- NVIDIA Blog
- OpenAI News
- Game Developer
- PocketGamer.biz
- AMD Newsroom
- 404 Media
- The Register
- MIT Technology Review
- Apple Newsroom
- Hugging Face Blog
- Microsoft Blog