NetherlandsBellingcat fed AI models 52 never-published photos to geolocate.
Anthropic claimed in September that its models approach superhuman geolocation, including for military targeting.
Opus, Fable and Google’s Gemini 4 Argon narrowed locations fastest, but none came close to perfect at the exact spot.
Bellingcat is developing agentic tools it says show promise for improving general-purpose agents’ geolocation.
How each outlet framed it
- Bellingcat
- tests unpublished photos to avoid training-data leakage, asks for one coordinate set, and includes open-weight Kimi and non-generative GeoClip
Sources: Bellingcat