War Map Live › How it works

How War Map Live works

War Map Live — often searched as warmaplive — is a live map of reported military incidents worldwide, assembled automatically from free public news feeds. No account, no paywall, and every marker links to the article it came from.

Who runs this

War Map Live is built and operated by Lumora Build. It is an independent project with no affiliation to any government, military, news organisation or political group, and it carries no advertising and no sponsorship — there is no ad unit anywhere on this site and nothing here is paid for. (It does load a Google Ads tag, which measures whether an ad brought you here; the privacy policy says exactly what that does.)

There is no newsroom behind it and no human reviews each incident before it appears — that is stated plainly because it is the most important thing to know about the data. What there is: a published method, a named operator, and a correction address. If a marker is wrong, tell us and it gets fixed: warmaplive@lumorabuild.com. Corrections are welcome and acted on; that is the only quality control this project has beyond the checks described below.

Corrections: warmaplive@lumorabuild.com — a wrong incident is corrected or removed.

1. Free sources only

Every five minutes the site checks each of its 72 public RSS feeds — Al Jazeera, BBC's regional desks, The Guardian, France 24, DW, Euronews, Sky News and the UN peace & security desk, plus outlets close to specific theatres: Ukrinform and Kyiv Post for Ukraine, The Jerusalem Post, Times of Israel, The New Arab, Anadolu, Middle East Eye and Al-Monitor for the Middle East, SCMP, The Hindu and Dawn for Asia, and Africanews and AllAfrica for Africa. No paid data, no API keys.

TASS, Tehran Times and IRNA are state-controlled and report claims as fact. They earn their place by covering incidents Western desks never touch, and the design carries the antidote: every marker names its outlet and links the original, so the attribution travels with the claim.

2. A filter, before any AI runs

Most of what these feeds publish is not an armed incident. A headline has to carry a concrete weapon or an armed actor and evidence that something actually happened before it reaches the model — so "fighting kills 12" passes and "peace talks resume" does not. Sport, business, crime and commentary are rejected outright. The filter is a set of word patterns, not an understanding of the story, so some off-topic headlines still get through.

3. The model classifies and locates — nothing else

A language model reads each surviving headline, decides whether it describes a specific armed incident at a specific place, and names that place, the type and the severity.

The model's coordinates are not trusted. Asked for a number, a model supplies one from memory, and that memory put the same Iranian island in the desert, off Qatar and in the Caspian Sea on three different days. So the place name is what the map believes: it is looked up in a checked list of the places conflict reporting actually names, and failing that in OpenStreetMap's geocoder (© OpenStreetMap contributors, ODbL). The model's own coordinate is kept only when it agrees with the name; a coordinate that is not even inside the country the report names is discarded rather than drawn.

It is never allowed to supply a source, an outlet name or a timestamp. Those are copied verbatim from the originating feed entry. Anything it returns that fails validation is discarded rather than repaired, and there is deliberately no seed or sample data — the map stays empty until something real lands on it.

4. An arc is only drawn when the article names an attacker

An arrow from one country to another is a factual claim, so one appears only when the reporting actually names who carried out the strike. Where nobody is named, the incident keeps a plain marker and no arrow. Even then the origin is an attribution, not an observed launch site.

Checks that run on every incident

None of these is a human review. They are the automated rules every incident passes before it is shown.

• Topic filter. Headlines are scored for weapons, armed actors and attack words before a model sees them, and commentary, crime and reaction pieces score zero. Some off-topic stories still get through; corrections are welcome.
• Location from the place name. The place is looked up in a checked gazetteer and then OpenStreetMap; a coordinate outside the country the report names is discarded.
• One marker per incident. Reports of the same type of incident in the same 0.1° grid cell (roughly 11 km) and the same three-hour UTC block are stored once.
• Attribution is copied, never generated. The outlet, the link and the time come from the feed entry.
• State media is labelled. Incidents from TASS, Tehran Times and IRNA carry a “state media” label on incident, country, archive and briefing pages.
• Translations are checked. A translation is rejected when its headline is still English, when a number differs from the English, or when it leaves out a summary that says more than the headline.
• Briefings are checked. A briefing is thrown away if it contains a number that is not in the incidents it covers or a count of them.
• Corrections. Write to warmaplive@lumorabuild.com and a wrong incident is corrected or removed.

What this is not

• Not verified. There is no reporter, no analyst and no human review between an outlet publishing something and it appearing here.
• Not precise. Coordinates are estimated from a sentence of text and are placed with far more apparent precision than the reporting supports.
• Not complete. An empty area means no monitored feed published a matching report — not that nothing happened. Coverage is uneven by region and by language.
• Not immediate. Timestamps are the article's publication time, which can be hours after the event.

Do not use it for safety, travel, operational, journalistic, insurance or legal decisions. The full disclaimer is in the Terms of Use.

Why the map defaults to six hours

Roughly three incidents an hour reach the map, and they arrive in bursts with a deep overnight lull. Measured across a full day, a third of the individual hours contained no more than one incident — so a one-hour view was frequently a world map with a single pin, telling the truth and looking broken. Six hours is the narrowest window that reliably has something in it. The window is absolute either way: nothing outside it is ever shown, and 1H is one tap away.