How we make the news
The Metro Map publishes short news items about metro and train networks. Most are drafted by an artificial intelligence model and published by automatic rules, without a person reading them first. This page explains how that works, what the rules check and what they cannot check. For how the map itself is built, see how we make the map.
Publishing principles
We publish summaries of what is happening on metro and train networks in the cities of our catalog, always with a link to the original article. We are not a team of reporters: we do not interview, investigate or take photos. Every item rests on an article published by someone else, which the text names and which is listed under Sources on the page. How we use those sources is in the sources policy. All items are in the news index.
Where the material comes from
Each city and network in the catalog has its own list of sources: RSS and Atom feeds from news outlets, registered by us. We read them regularly and keep, for each item, the headline, a short summary, the link and the date. We never keep the full text of the original article.
Before an item is used for an article, it goes through an automatic selection. We discard it when:
- it is not clearly about metro or train services (a score based on words in the language of the source);
- it is more than 30 days old when collected;
- it is a duplicate: the same address, or a headline at least 75% similar in the same city within 14 days;
- it is about another city or network of the catalog, not the one of the source.
Artificial intelligence
The texts are drafted by Claude Haiku, an artificial intelligence model from Anthropic. For each selected item, our server downloads the original article and gives its text to the model, then discards that text after the draft is written. The model writes, in English and in its own words, a title, a description, key points and the body, naming the source. It also tells us which city it believes the text is about and how confident it is.
The Portuguese (pt-BR) version of an article is a translation of the English one made by an AI model and checked by its own rules: numbers, links, the source named, language and length must match the original. Models can make mistakes, which is why the automatic checks below exist and why every article links to its source, so you can compare.
Verification and automatic checks
No fact-checker reads the text. Instead, software compares it with the original article and holds the text back if any of these rules fails:
- the full text of the original article was available, not just the summary from the feed;
- the body has between 150 and 600 words, and a title and description;
- the text names the outlet of the original article, and there is a link to it;
- the text is in the expected language;
- it is not too close to the original: at most 35% of five-word sequences identical, and no run of 20 or more identical words;
- the numbers in the text appear in the original (compared by value), except years, numeric dates (such as 14/10) and line numbers, which this rule does not check; and at least 3 facts (numbers and names) of the original are kept;
- every line and station it mentions exists in the network's catalog or in the original;
- the model confirmed that the text is about the city and network of the item, the city it reported is the right one, and the text does not mention other cities or networks of the catalog;
- the model's own confidence is at least 0.7;
- it does not touch a sensitive topic: deaths, injuries, attacks, bombs, shootings, derailments, collisions, evacuations, victims and similar words, in several languages. These are never published automatically;
- the body links only to this site and to the listed sources;
- it is not a duplicate of an article already published (the same item, or a very similar title in the same network within 14 days).
A text that fails any rule is not published: it stays as a draft that visitors cannot see, with the reasons recorded. These rules compare the text with its source; they do not check the source against the world, so they reduce the risk of errors but do not remove it.
No human review before publishing
No person reads a text before an article goes live. When a text passes all the rules it is published and offered to search engines straight away. That is why every article produced this way carries this notice:
Text produced with the help of artificial intelligence from the articles listed under Sources. Checked by automated rules, without human editorial review.
Daily re-check
The checks do not stop at publication. The re-check runs once a day over the AI articles already published, applying again the rules that work on the saved text: city and network, length, the source named, language, sensitive topics, the link to the source and outside links. An article that no longer passes goes back to draft and disappears from the site, together with its translation. A draft that had been held back only for language or attribution can go live when it passes the current rules.
Corrections
If an article is wrong, tell us on the contact page, with the address of the article and what is wrong. We look at every report and correct the text or remove the article. When a published article is changed more than ten minutes after publication, its page adds the time of the update next to the time of publication; smaller adjustments made right after publication are not shown. Rights holders and people mentioned in an article can ask for removal through the same channel.
Feedback
Questions, complaints or suggestions about the news are welcome on the contact page. Tell us which article (its address) you mean and what you would like us to know.
Why articles have no individual author
Our articles carry no author's name because no individual author writes them: they are drafted by an AI model from the sources and published by automatic rules. We do not invent authors. The articles are credited to The Metro Map as publisher, which is also the author in the page's structured data, and the original reporting is credited to the outlet named in the text and linked under Sources.
Related: how we make the map, the sources policy, the news index and the about page.