Use of artificial intelligence

1. What artificial intelligence generates

Our own headline and summary of each story ("What the source reports"), in Spanish and English, from the headline and text published by the outlet. The summary is limited to the basic fact (what, who, where, when), without quotes or the outlet's analysis. They are not translations: the AI writes in its own words and with its own structure.

The "AI interpretation": the impact analysis and, for stories about wars, sanctions or treaties, the possible economic impact. These are inferences, not verified facts.

The tone classification ("Hostile", "Beneficial", "Neutral" or "Irrelevant") and the countries or companies involved, which draw the connections on the globe.

The question shown above the headline, written for search engines.

If you choose French or German, the translation of the above. And the comparison of narratives between two countries.

2. What it does NOT generate

Links, dates and outlet names, which come from the source itself.

The analyses signed by the owner ("Analysis" section). If one is drafted with AI assistance, the analysis says so.

3. How each story shows it

Each story carries a visible notice: "Generated by AI, not reviewed by a person" or "Generated with AI and reviewed by the editor", with the generation date, next to the "Report an error" link.

No story is published before it has been analysed: until then it waits in a queue.

Stories that attribute crimes or wrongdoing to identifiable people, or concern victims or minors, are not published until a person has reviewed them.

"Reviewed by the editor" only appears when a person has checked the facts against the source and takes editorial responsibility.

The page data also tells search engines and other machines that the text was generated by AI.

4. Checks and human review

Writing rules: neutral news-agency tone; each party's claim is attributed to whoever makes it; accusations are presented as accusations; private individuals without public office are not named; no war propaganda and no investment recommendations.

Automatic checks: sentences that sound like buy or sell advice or price forecasts are removed; we measure how similar the headline and summary are to the original and, if too similar, the AI rewrites them and, if it fails, the story is not published; and stories about deaths, attacks, terrorism, crime, corruption or sanctions go to a human review queue.

These checks are technical: they reduce risk but do not guarantee on their own that there are no errors.

The owner reviews that queue, corrects or removes what does not comply and marks reviewed stories. Every correction keeps the previous version.

Traceability: for each story we store which model wrote it, with which version of the rules and when.

5. Limitations

Language models make mistakes: they may misjudge the tone, attribute an event to the wrong country or summarise inaccurately. Except for those that affect identifiable people, we do not review every story by hand before publishing it.

No label on this site is an editorial judgement or a verified fact. If in doubt, read the story on the original outlet.

Analyses age: stories older than a week and analyses older than three months say so.

6. Models and data

We use Google's Gemini models through their API. Only the public text of news stories is sent to them, and it is deleted from our systems after the analysis. No data about visitors or from the forms is sent.

7. How to challenge a result

Click "Report an error" on the story or use the contact form (or write to contacto@geostock3d.com). We review every request; if there is an error, we correct or remove the story and keep a record of the correction.