How Bellingcat Uses AI and Satellite Imagery in Open-Source Investigations

Updated

Bellingcat combines AI tools with satellite imagery as aids to human verification, not replacements for it. Its published work shows LLMs being tested to speed up geolocation, while satellite images are used to confirm or challenge viral claims about events such as the Gaza ruins, new border road construction in Big Bend National Park, and tree-felling at a DC golf course. The consistent condition across these cases: machine output is treated as a lead that a human researcher then checks against primary imagery and other sources.

AI in the Investigation Workflow

Bellingcat's most direct statement on AI comes from its geolocation testing: LLMs can "supercharge" geolocations, with tests showing big improvements — but humans remain key. In practice, that means AI is used to generate candidate locations, translate or summarize local sources, and narrow the search space before a researcher confirms a match.

A second documented use is detecting manipulated media. In its coverage of the India-Pakistan conflict, Bellingcat examined how a deepfake video moved into the mainstream — an example of AI-assisted analysis applied to verifying whether footage is authentic rather than to locating it.

The workflow implied by these examples:

  1. Input — a viral image, video, or claim with unclear origins.
  2. AI-assisted triage — candidate locations, translations, or manipulation flags.
  3. Human verification — cross-checking against satellite imagery, maps, and primary sources.
  4. Output — a published finding that states what the imagery does and does not prove.

Satellite Imagery as the Verification Layer

Satellite imagery is where Bellingcat's human-checking step becomes concrete. Documented examples from its investigations:

Investigation What imagery established
Gaza ruins More recent satellite imagery obtained after viral Google Maps images circulated, giving a fuller picture than the widely shared posts
Big Bend National Park Satellite imagery documenting new Border Patrol road construction
DC golf course Imagery showing the extent of tree-felling by the Trump administration
Mali drone campaign Geolocation and imagery supporting a pattern of civilian harm

The Gaza case illustrates the core method: a viral social media image is treated as a claim, then compared against newer or higher-resolution imagery to see what changed and what the post left out.

Guides for Learning the Techniques

Bellingcat publishes step-by-step guides alongside its investigations, including "Burning Forests: Tools for Tracking and Reporting Wildfire Damage" and "How to Use AI to Help Find Civilian Harm." These are the practical entry point for anyone wanting to replicate the AI-plus-imagery workflow. The site also runs a membership program offering advanced access to training announcements and podcasts where researchers discuss methods behind specific stories.

What This Means If You Want to Apply the Approach

  • Start with a verifiable claim. Pick an image or video with a stated or implied location and time.
  • Use AI for speed, not conclusions. Let it suggest candidate sites or flag possible manipulation, then verify each one yourself.
  • Go to primary imagery. Compare the claim against satellite or map imagery that is newer or clearer than what circulated.
  • Report the limits. State what the imagery confirms and what remains unresolved — the standard Bellingcat applies to its own published work.

The recurring lesson from Bellingcat's own testing is that AI shortens the search but does not remove the need for a human to confirm the match against source imagery.