Skip to content
Menu
Report an incident DE : Deutsche Version
Automatic processing

Automatic processing

Which details on the video pages are generated with AI assistance from the original description, what stays unchanged, who checks and how to report errors.

Last updated: Thursday, 8 October 2026

What is generated automatically

For more than 4,300 videos the details of the video pages are not written by hand. A language model reads the original video description on YouTube and produces from it:

  • a readable title and a short summary (the descriptions on YouTube contain capitals, emojis and abbreviations),
  • the incident type (fire, traffic accident, technical rescue, hazardous materials, rescue, police operation, exercise and others) and the services (fire brigade, police, THW, EMS, emergency vehicles, exercise, trade fair),
  • the location (place, district, state), from which the map position is derived as well,
  • the meta description for search engines and the English version of title and summary.

The pages are labelled accordingly: below the facts there is a note that title, teaser, category and location were created with AI assistance. Once a person has checked and set the title of a page, the note disappears.

What is not generated automatically

  • The authority’s press release is taken from the video description unchanged. It is not rephrased, shortened or extended.
  • The model may add no facts: the number of injured, the severity or a fatality are adopted only when the text states them explicitly; otherwise these fields stay empty.
  • Cause and blame are never attributed. Nationality or origin of persons appear neither in titles nor in summaries or tags.
  • Before a text goes to the model, age statements with origin, house numbers, number plates and phone numbers are removed.

Model, provider and control

Processing runs through the OpenRouter service (USA) with a language model selected there; which model is currently in use is stated in the privacy policy. Only endpoints without data retention are used.

The result is not published blindly: details with low model confidence land on a review list, samples are checked monthly, and every detail can be corrected by hand in the editorial system. A manual correction survives every later automatic run. Editorial responsibility for the pages lies with Shooter-Media-TV, see the imprint.

Report an error

A wrong place, a wrong incident type or a misleading title? Please report it through the contact form with the address of the page. Corrections are usually made within 24 hours.

This labelling follows Article 50(4) of the AI Act, which obliges deployers to disclose when published texts on matters of public interest were artificially generated or manipulated. Which data is processed in the course of it is explained in the privacy policy.