The challenges of conducting open-source research in China are well-documented. Consistently named one of the most digitally oppressive countries in the world, China blocks some of the world’s largest social media platforms, such as Facebook, Google, and YouTube. Those that are still accessible are mostly Chinese-owned, strictly regulated and monitored in real time by AI systems as well as tens of thousands of “internet police”. 

But despite these strict controls, Chinese apps – which boast more than a billion estimated users – remain an information goldmine for investigative journalists covering stories both within and outside China.

Support Bellingcat

Your donations directly contribute to our ability to publish groundbreaking investigations and uncover wrongdoing around the world.

Donate

Since most foreign sites are banned, Chinese platforms are the largest resource available to journalists and researchers interested in what’s going on in the world’s second-most populous country. Even when a topic is being censored, patterns in the censorship can themselves serve as investigative leads: a 2020 BuzzFeed News investigation, for example, mapped out detention camps in Xinjiang by examining areas that had been blanked out on China’s Baidu Maps.

With millions of Chinese people living overseas, social media activity by members of the diaspora can also turn into global stories.

Serial rapist Zou Zhenhao, a Chinese PhD student, was jailed in London last year after one of his victims posted a warning on Xiaohongshu, also known as Little Red Book or Rednote, an app popular with young Chinese women living abroad. Another woman Zou had raped reached out to the original poster, who put her in touch with the police – leading to the conviction of a man described by police as possibly one of the worst sexual predators in British history.

Founded in 2013 as a Hong Kong shopping guide, Xiaohongshu has evolved into a lifestyle and e-commerce platform that has been compared with Instagram, Pinterest and Amazon. Last year, it reported about 300 million monthly active users, rivalling some of China’s largest social media platforms.

The app’s 600 million daily searches by the end of 2024 also accounted for half of market leader Baidu’s search volume, demonstrating that it is emerging as a critical search and discovery engine, not just a social platform.

Although primarily a Chinese-language app, Xiaohongshu gained attention in the English-speaking world last year, when millions of American TikTok users flocked to the platform in anticipation of a TikTok ban under US President Donald Trump. 

Responding to the surge of international users – sparked by the #TikTokRefugees trend – Xiaohongshu rolled out an AI-powered translation feature, making the app more accessible to non-Chinese audiences. This also meant that journalists without Chinese language skills can more easily communicate on and navigate the platform.

Despite its growing popularity both within and outside China, the app is relatively new and underexplored compared to more well-established platforms such as Weibo. 

This guide aims to provide a starting point for those looking to explore Xiaohongshu for open-source investigations, including an overview of its main user demographics, potential topics to explore and strategic search methods specific to the app. 

User Demographics and Topics

According to Xiaohongshu’s official data, the platform’s demographic profile is mainly young, female and urban. As of 2024, 70 percent of its users were women, with half of all users belonging to Gen Z and living in China’s largest cities. 

As previously mentioned, the app has also gained popularity with the Chinese diaspora. Many Chinese nationals living abroad use it as a search engine for local information, posting and searching for content related to their daily lives, from restaurant recommendations and apartment hunting to navigating foreign bureaucracies and finding community resources. 

This demographic profile makes Xiaohongshu particularly well-suited for investigating stories about consumer fraud and urban livability issues. For example, Chinese outlets like Jiemian have used Xiaohongshu posts to expose the grey-market ecosystem of paid reviews and fake endorsements tied to the platform’s e-commerce model, while in 2022, International Financial News traced a mother-and-baby store scam that defrauded over 400 parents back to product recommendation posts on the platform.

Given its predominantly female user base, Xiaohongshu has also evolved into one of China’s most important spaces for feminist discourse and women’s issues. Academic researchers have used content on the platform to analyse local discussions on menstrual shaming, sexual harassment, and the controversial “divorce cooling-off period” introduced in 2021. As Rest of World reported, women have increasingly congregated on Xiaohongshu, where they outnumber male users and have found ways to trick the app’s recommendation algorithm so their posts are shown mostly to other women.

The Relevance of Censorship

Political content and current affairs about China are largely absent from the app – a result of both active censorship and platform design. 

All Chinese social media platforms, including Xiaohongshu, operate under strict content moderation requirements from the Cyberspace Administration of China. A leaked 143-page internal document published by China Digital Times in 2022 revealed how Xiaohongshu censors respond to government directives in “real-time”, blocking content related to politically sensitive topics such as criticism of the Chinese Communist Party, labour strikes and student suicides. Xiaohongshu’s commercial focus also makes it less likely that these topics would be discussed on the platform: as Rest of World reported, the platform functions less like Weibo – a public square for current events – and more like “a giant mall, where shoppers tell each other what to buy”.

Related articles by Bellingcat

Resources

The Challenges of Conducting Open Source Research on China

Coverage of international affairs is also tightly controlled: only state-owned or state-controlled news organisations can obtain licences to publish original news content. However, content about life abroad, particularly stories about the cost of living, healthcare, or social problems in Western countries, circulates more freely on platforms including Xiaohongshu, and provide journalists with insight into how Chinese diaspora communities engage with local political systems. 

For example, when the 2025 Miss Finland was accused of making anti-Asian gestures, searching for “芬兰小姐” (Miss Finland) and “投诉” (complaint) on Xiaohongshu revealed a trove of collective action: users shared different complaint pathways, posted templates for filing reports, and documented various outcomes from their complaints. 

For such large-scale public events, Xiaohongshu can be both an organising platform and a rich source for tracking how diaspora communities coordinate responses to discrimination, providing journalists with insight into grassroots activism and transnational advocacy networks.

Getting Started

Xiaohongshu is available for download on both Apple’s App Store and Google Play worldwide, or can be accessed via a web browser. In international app stores, the app appears under the name “RedNote,” but this is the same application as Xiaohongshu – content and accounts are shared across both. The key difference is that RedNote users who register with overseas phone numbers are automatically tagged as international users, which affects the content the algorithm surfaces to them.

For users who download the app outside mainland China, Xiaohongshu automatically detects the device language and location. Upon first login, international users are prompted with an option to automatically translate all content into English (or their device language). If enabled, posts and comments will display with translations by default, and the algorithm will prioritise English-language content and posts created by or for international users, such as expat influencers.

For researchers and journalists seeking to observe the platform as Chinese users experience it, consider disabling automatic translation. This allows you to see content as it natively appears and helps you distinguish between posts created for international audiences versus those created for domestic users – a distinction that matters when assessing how representative your sample is for the relevant topic.

The default home feed, or the “Explore” tab, is where the algorithm surfaces content based on your engagement history, location and user profile. The feed uses a grid layout displaying post thumbnails with titles and like counts.

On the top right corner of the screen, the search bar also allows keyword searches across posts, users and topics. Results can be filtered by content type (e.g. notes, videos, users or products) and sorted by relevance or recency.

Using the Search Bar

Xiaohongshu’s search function is relatively basic. You can search by keywords and filter by time and location, but the options are general: time filters include “past day,” “past week,” or “past six months,” while location filters offer “same city” or “nearby”. 

For example, searching “Canada” returns posts tagged with that keyword, which you can then sort by recency or proximity. 

For breaking news events, try searching location names or names of individuals involved in the incident, filtering for the most recent posts to capture real-time reactions and on-the-ground accounts before they’re censored or deleted.

Xiaohongshu primarily uses algorithms to curate and push content through personalised feeds. For journalists using Xiaohongshu for investigative purposes, it can be useful to actively search for topics of interest to train your algorithm – the more you search and engage with specific content, the more relevant posts the algorithm will surface to you.

However, if you are researching the platform itself – studying what content Xiaohongshu promotes, how censorship operates, or what narratives dominate – you may want to start from a clean slate. In that case, consider periodically turning off personalised recommendations (Settings → Privacy Settings → Personalisation Options), clearing your browsing history, clearing cached data, or using a fresh account to observe what the platform shows to a “neutral” user.

Language and Lingo

During the influx of “TikTok refugees” in January 2025, Xiaohongshu launched a translation feature for users outside mainland China, enabling the automatic translation of comments and posts. 

However, this does not translate search queries. The platform’s search engine is still optimised for Chinese, though there is a “prioritise English” filter for overseas users, and searching in English will return some results.

But the language you search in shapes far more than just your results – it determines which version of the platform you see. When you search in English or use an international account, the algorithm treats you as a foreign user and surfaces content accordingly: influencers explaining why they love living in China, comparisons showing Chinese life favourably against the West. 

This isn’t a neutral cross-section of the platform – it is a curated bubble. To access what Chinese users actually discuss among themselves, it would be more effective to search in simplified Chinese and, ideally, use a China-registered account if you have access to one. If you don’t read Chinese, you can also consider using a translation tool (Google Translate, DeepL, or an AI assistant) to convert your search terms into simplified Chinese before entering them.

Despite such tools and the in-app translation feature, it is always useful when researching using Chinese platforms to work with a native speaker familiar with the local context. They can flag when an innocuous-seeming term actually carries hidden meaning, and help identify coded conversations about a censored topic.

On Xiaohongshu specifically, this coded language extends beyond political topics to include anything the platform’s algorithm might flag as “vulgar” or promotional. For example, users substitute fruits and neutral terms for body parts or sexual content to avoid being flagged as inappropriate – the peach emoji for buttocks, or 炒菜 (“cooking”) for explicit material. They may also use abbreviations and emojis for commercial terms to evade anti-marketing filters, such as “vx” (the abbreviation of how WeChat is pronounced in Chinese) or “绿” (“plus green”, apparently referring to WeChat’s green logo) for WeChat, or “米” (rice) or the moneybag emoji for money.

Advanced Search Strategies

For more sophisticated searching, consider using third-party marketing analytics tools like Xinhong and Qiangu, which can show trending topics, popular posts and engagement metrics, as well as identify key content creators posting about specific subjects. 

For example, on Xinhong, when you search for “Canada” in Chinese, it also shows show trending related searches such as “加拿大总理” (Canadian Prime Minister). Clicking through these suggestions leads to recent posts—for example, posts about Mark Carney’s latest statements at Davos, along with user comments and reactions.

While these tools are designed for marketers, they provide journalists with valuable capabilities: tracking how topics evolve, identifying influential voices in specific communities, and discovering related hashtags or discussions that might not surface through basic platform search. These tools often require paid subscriptions but can significantly enhance research efficiency for long-term investigations.

Another valuable feature is Xiaohongshu’s group chat function, where users gather around shared keywords and topics—from city-specific communities to niche interests. These groups are often highly active and provide access to candid community discussions that don’t appear in public posts. To find relevant groups, go to Messages → Group Square, where you can browse categories or search by keyword and request to join.

Monitoring active group chats related to relevant topics, whether that’s a specific city, industry, or issue, can help journalists and researchers stay updated on emerging issues and detect potential story leads before they become widely visible on public feeds.

Preserving the Evidence

Chinese social media content can disappear quickly and without warning due to censorship, making immediate preservation critical. 

Always take two preservation steps immediately upon discovering relevant content:

First, screenshot the entire post, including the URL, timestamp, username, like/comment counts, and location tags. These metrics establish context and authenticity. Use tools that capture full-page screenshots rather than just visible portions, as posts can be long and comments extensive. Second, archive the web page using services like archive.today or Wayback Machine. Note that these services capture only static content – comments and engagement metrics may not be fully preserved and should be screenshotted separately.

For Xiaohongshu specifically, always preserve the user’s unique ID found in their profile URL when viewed on a browser, which follows the format “user/profile/[unique ID]”. Users can change their display names, but this unique identifier remains constant, allowing you to track accounts over time even after name changes. This is critical for long-term investigations or when monitoring specific sources.

Xiaohongshu operates under the same legal and censorship constraints as all Chinese social media platforms, and researchers should approach it with appropriate caution. Content moderation is extensive: users who post about sensitive subjects risk having their content removed or their accounts suspended, and the platform is required to comply with government data requests. For researchers, this means the information you find represents only what has survived the censorship process.

That said, Xiaohongshu remains a remarkably rich resource for open-source research. Its strength lies precisely in its apolitical, lifestyle-oriented identity: while political discussion is suppressed, candid conversations about everyday life flourish. For journalists willing to invest in learning the platform’s rhythms, building Chinese-language search skills, and understanding its coded vocabularies, Xiaohongshu offers a window into how ordinary Chinese people talk among themselves – an area that remains largely untapped by international media.

Bellingcat is a non-profit and the ability to carry out our work is dependent on the kind support of individual donors. If you would like to support our work, you can do so here. You can also subscribe to our Patreon channel here. Subscribe to our Newsletter and follow us on Bluesky here, Instagram here, Reddit here and YouTube here.

The post Mining China’s ‘Little Red Book’ for Open Source Gold appeared first on bellingcat.

Access to open source visuals of the current Iran conflict, which has spread to many parts of the Middle East, continues to be sporadic. Videos and photos from within Iran trickle out on social media as the Iranian internet blackout hinders the flow of digital communication. 

In past conflicts, satellite imagery has provided a vital overview of potential damage to both military and civilian infrastructure, especially when there are digital black spots or obstacles to on-the-ground reporting. But imagery from commercial providers is becoming increasingly restricted, leaving even those who have access to the most expensive imagery in the dark. 

Shortly after the war in Gaza began in 2023, Bellingcat introduced a free tool authored by University College London lecturer and Bellingcat contributor, Ollie Ballinger, that was able to estimate the number of damaged buildings in a given area. This helped monitor and map the scale of destruction across the territory as Israel’s military operation progressed. 

Bellingcat is now introducing an updated version of the open source tool — called the Iran Conflict Damage Proxy Map — focused on destruction in Iran and the wider Gulf region. 

It can be accessed here.

How it Works

The tool works by conducting a statistical test on Synthetic Aperture Radar (SAR) imagery captured by the Sentinel-1 satellite which is part of the Copernicus mission developed and operated by the European Space Agency. SAR sends pulses of microwaves at the earth’s surface and uses their echo to capture textural information about what it detects. 

The SAR data for the geographic area covered by the tool is put through the Pixel-Wise T-Test (PWTT) damage detection algorithm, which was also developed by Ollie Ballinger. It takes a reference period of one year’s worth of SAR imagery before the onset of the war and calculates a “normal” range within which 99% of the observations fall. It then conducts the same process for imagery in an inference period following the onset of the war, and compares it to the reference period. The core idea is that if a building has become damaged since the beginning of the war, then the “echo” (called backscatter) from that pixel will be consistently outside of the normal range of values for that particular area. Investigators can then further probe potential damage around this highlighted area.

The plot below shows how the process was applied to Gaza and several Syrian, Iraqi and Ukrainian cities. The bars represent the weekly total number of clashes in each place, sourced from the Armed Conflict Location Event (ACLED) dataset. The pre-war reference periods are shaded in blue, spanning one year before the onset of each conflict. The one month inference periods after the respective conflicts  began are shaded in orange. The blue and orange areas are what the tool compares. 

The plot below shows an area with a number of warehouses in Tehran’s southwest. Some of the buildings show clear damage in optical Sentinel-2 imagery (something that has to be accessed outside of the tool via the Copernicus Browser). 

Clicking on the map within the tool generates a chart displaying that pixel’s historical backscatter; the red dotted lines denote a range within which 99% of the pre-war backscatter values fall. In this example, we can see that from March 14 onwards, the backscatter values over this warehouse begin to consistently fall outside of their historical normal range. This could signal that damage has been detected in the area.

Two important aspects of this workflow are that it utilises free and fully open access satellite data, as opposed to commercial satellite services; the second is that it overcomes some key limitations of AI in this domain, the most serious of which is called overfitting. This is where a model trained in one area is deployed in a new unseen area, and fails to generalise. Because we’re only ever comparing each pixel against its own historical baseline, we don’t run into that problem. 

Accuracy

The PWTT has been published in a scientific journal after two years of review.  Its accuracy was  assessed using an original dataset of over two million building footprints labeled by the United Nations, spanning 30 cities across Gaza, Ukraine, Sudan, Syria, and Iraq. Despite being simple and lightweight, the algorithm has been recorded achieving building-level accuracy statistics (AUC=0.87 in the full sample) rivaling state of the art methods that use deep learning and high resolution imagery. The plot below compares building-level predictions from the PWTT against the UN damage annotations in Hostomel, Ukraine. True positives (PWTT and United Nations agree on damage) are shown in red, true negatives are shown in green, false positives in orange, and false negatives in purple. The graphic shows the accuracy of the tool, while also emphasising that further checks on what it highlights should be conducted to draw full conclusions.  

It is important to note that just because the tool may show a high probability of a building or buildings being damaged or destroyed, that doesn’t make it definite. 

It is best to check with any other available imagery — either open source photos and videos that’ve been geolocated by a group such as Geoconfirmed or Sentinel-2 as well as other commercial satellite imagery if it’s up-to-date for the area. At time of publication, Sentinel-2 satellite imagery still offers coverage over the area that the tool focuses on. Other commercial satellite imagery providers have limited their coverage.

What the tool excels at is highlighting and narrowing down areas so that further corroboration or further confirmation can be sought.

Testing the Tool

Using the Iran Conflict Damage Proxy Map, we can spot some of the larger areas of potential damage or destruction that have occurred since the Iran war started. 

Starting from a zoomed-out view of Tehran, there are a few spots that appear with large clusters of high damage probability. Cross-referencing these locations with open source map data from platforms like OpenStreetMap or Wikimapia, we can start finding sites that would make for likely targets – such as military sites.

One example of a potentially damaged site visible in the tool is the Valiasr Barracks in central Tehran, which was struck in the first week of the war. By going to the Copernicus Browser and reviewing the area with optical Sentinel-2 imagery, we can see clear indications of damage at the barracks.

IRGC Valiasr Barracks in Tehran:

A large Islamic Revolutionary Guard Corps (IRGC) compound near Isfahan is another example of military infrastructure that is readily visible in both the Iran Conflict Damage Proxy Map as well as Sentinel-2 imagery. 

IRGC Ashura Garrison in Isfahan:

Air bases have also been a frequent target for U.S.-Israeli strikes in Iran. The Fath Air Base just outside of Tehran, near the city of Karaj, shows the signature of potential damage when using the tool. Checking Sentinel-2 imagery shows damage to multiple large buildings on the northern side of the base.

Fath Air Base in Karaj:

The U.S. has stated that destroying Iran’s “defense industrial base” is also a goal, which makes large areas like the Khojir missile production complex east of Tehran a good location to search with this tool. The tool suggests large clusters of damage on both the eastern and western sides of the complex — near areas where solid propellant is reportedly produced and where other fuel components are reportedly made.

Khojir Missile Production Complex outside of Tehran:

Usage in the Gulf Region

While useful for providing a sense of damaged areas in Iran, the Iran Conflict Damage Proxy Map can also be used to see damage outside of Iran, particularly at sites in the region which Iran has been targeting with drones and missiles.

In the below example at Al Udeid Air Base in Qatar, which hosts U.S. Central Command’s Combined Air Operations Center, there is a notable indication of damage over a warehouse-like building at 25.115647, 51.333125. Checking the same location in Sentinel-2 imagery shows that there does appear to be damage at that warehouse — represented by a large blackened area on the white roof. According to Qatar’s Ministry of Defense, at least one Iranian ballistic missile struck the base in early March.

Al Udeid Air Base in Qatar:

Civilian sites struck by Iranian drones or missiles are also visible in the tool — though the damage has to be fairly large in order to be picked up. Something like damage to the sides of high rise buildings from an Iranian drone attack doesn’t readily appear in the tool. Sites that do appear are places like oil refineries, such as a fuel tank at Fujairah port in the United Arab Emirates. 

Fuel tanks at Fujairah Port, UAE:

Accessing the Tool

It’s important to keep in mind that the data for the Iran Conflict Damage Proxy Map is updated approximately one or two times per week as new satellite data is collected by the Sentinel-1 satellite, so it’s not meant to be a representation of real-time damage to buildings. 

Still, it can be useful for researchers to quickly gain an overview of damage throughout Iran and the Gulf where suspected strikes may have taken place and when there is no other open source information available.

You can access the Iran Conflict Damage Proxy Map here.

Similar tools using the same methodology to assess damage in Ukraine following Russia’s full-scale invasion and Turkey following the 2023 earthquake can be found here. The Gaza Damage Proxy Map can be found here. 

Bellingcat’s Logan Williams contributed to this report.

This article was updated on April 7, 2026, to note that Sentinel-1 and Sentinel-2 are part of the Copernicus mission developed and operated by the European Space Agency.

Bellingcat is a non-profit and the ability to carry out our work is dependent on the kind support of individual donors. If you would like to support our work, you can do so here. You can also subscribe to our Patreon channel here. Subscribe to our Newsletter and follow us on Bluesky here, Instagram here, Reddit here and YouTube here.

The post When Satellite Imagery Goes Dark: New Tool Shows Damage in Iran and the Gulf appeared first on bellingcat.

Since launching the military campaign against Iran on Feb. 28, the US and Israel have dropped thousands of bombs on the country. Videos of explosions have become a source of misinformation and misunderstanding, with many of the strikes incorrectly attributed to a particular munition and many explosive effects – seen in footage and images – falsely attributed to “mystery” or illegal weapons.

Take the below post that initially suggested (although it said more analysis was required) that the US may have used a nuclear weapon in Iran, an outlandish and clearly incorrect claim that experts Bellingcat spoke to had little time for.

IMPORTANT UPDATE AND NOTE: The following is not a complete assessment and I require more data to verify first use. This is a surface level observation but it must be noted.

The US used what appears to be, without additional details, a nuclear weapon on Iran delivered by a… pic.twitter.com/7ucJNdGyNi

— Korobochka (コロボ) (@cirnosad) March 11, 2026

The post, set to private after the publication of this guide, appeared to suggest that a nuclear explosion happened in Iran. Source: X/cirnosad

“The video does not show a nuclear explosion—something that I am astonished even needs to be clarified,” Dr NR Jenzen-Jones, Director of Armament Research Services, a weapons intelligence consultancy, told Bellingcat.

Mushroom clouds can form when explosions produce hot gases that quickly rise and encounter resistance from denser, colder air. (Clouds created by nuclear weapons can also vary significantly in appearance.)

“Certain types of explosive munitions, such as those working on the fuel-air explosive (FAE) and thermobaric principles, are particularly poorly understood by non-specialists. As a result, these and other types of munitions are routinely misidentified,” Jenzen-Jones said.

Support Bellingcat

Your donations directly contribute to our ability to publish groundbreaking investigations and uncover wrongdoing around the world.

Donate

Often posts about explosives are incorrect or inaccurate because of a lack of knowledge about how explosives work, but in other cases misinterpretations are deliberate. Joe Dyke, director of programmes at Airwars, told Bellingcat that deliberate disinformation that shifts responsibility of a strike is the most common they see, with posts often sharing flimsy but “scientific sounding” analysis.

Better understanding explosives can make it easier to identify misinformation surrounding explosions. 

This guide explains explosives, their characteristics and the impact they have on people and infrastructure. We highlight the differences between thermobaric and Dense Inert Metal Explosives (DIME), two types of explosives that are frequently the subject of misinformation.

What Are Explosives?

Explosives are energetic materials capable of causing death and destruction through a rapid release of energy. The blast creates pressure waves emanating from the epicentre. These waves can directly kill or injure people and shatter objects into lethal fragments.

High explosives are typically used in warheads and shells; they differ from low explosives which are often used in rocket propellants. The supersonic speed of the explosive reaction- classified as detonation- also separates the two kinds of explosives. During detonation, temperatures can rise above 3,000 °C, but only briefly and very close to the reaction zone, Dr Sabrina Wahler, a Postdoctoral Scholar at the California Institute of Technology focusing on research of detonation products told Bellingcat.

The detonation creates a shockwave, which is a visible wave or bubble in high speed videos. The shockwave impacts people and objects before the sound of the blast can be heard.

The shockwave is the result of the pressure pushing air away from the blast in the positive phase. When the air rushes back in the negative phase, it creates a suction effect.

The shockwave arrival time, combined with a known distance, has been used to estimate the explosive weight of blasts, including the Beirut explosion in 2020.

Reactive materials, such as aluminium powder, are often added to explosives to improve performance. These metals react with the gaseous products from the detonation, resulting in increased energy output, Jacqueline Akhavan, a Professor of Explosive Chemistry at Cranfield University, told Bellingcat.

Sometimes, reactive metals such as aluminium from the explosive composition can be seen burning outside the fireball, indicating an explosive with reactive metal.

The size of a fireball does not necessarily indicate the blast’s power. In movies and airshows, a “Hollywood shot” involves igniting large amounts of gasoline with small amounts of explosives, creating spectacular fireballs with minimal pressure.

Thermobaric, and dense inert metal explosives (DIME), are other types of explosive compositions where metals are added to modify specific effects.

Thermobaric Explosives

In January 2024, after an attack in Gaza, social media posts appeared claiming that thermobaric explosives “literally sucks the air out of the children’s lungs and causes them to internally explode”. According to an article by Dr Rachel Lance, a biomedical engineer specialising in patterns of injury and trauma from explosions “there is no evidence that thermobarics pull the air out of the lungs”. 

There were also claims that thermobaric weapons incinerate people. According to a report by the Armament Research Services, the effects of this type of explosion “are of the same nature as those expected from a conventional high explosive”. The only difference is that the duration of each effect is likely to be longer from a few milliseconds to tens of milliseconds and in a pressure wave with a lower peak.

This occurs because thermobaric explosives add a significant amount of fuel or reactive metals to the explosive composition. Some of the fuel burns after detonation. These munitions are effective against cave or bunker systems, as the pressure wave can travel further throughout the structure.

Visual differences can indicate the types of explosives used. Even within the same category, explosives may appear different because of variations in chemical composition, conditions where the explosion occurs, and video quality.

TÜBİTAK SAGE’den yerli termobarik patlayıcıda yeni bir adım daha!

Kapalı alanlarda yüksek darbe ve sıcaklık etkinliğine sahip yeni bir termobarik patlayıcı

TENDÜREK’ten sonra KOR ile geleneksel patlayıcılara göre 4 kat daha yüksek sıcaklık etkinliği pic.twitter.com/N4yZ8YvMi9

— TÜBİTAK SAGE (@SageTubitak) March 5, 2020

Comparison of KOR, a thermobaric explosive, and TNT, in a test by TÜBİTAK SAGE, a Turkish Defense Research Organization. Source: X/TÜBİTAK SAGE.

Many countries, including the US, Russia, China, Ukraine, Iran and Turkey, use enhanced blast and thermobaric explosives. Russia has used them in Ukraine and Syria. Israel uses munitions that have variants featuring thermobaric warheads, but the use of thermobaric explosives has not been confirmed.

Fuel-air explosives are similar to thermobaric explosives, but function differently. Both are volumetric weapons, but fuel-air explosives disperse a cloud of fuel, then the explosion occurs.

Dense Inert Metal Explosives

Unsubstantiated claims of DIME munitions have regularly surfaced since 2006, when they were first alleged to have been used in Gaza. Similar claims have reappeared in Gaza since the war began on Oct. 7, 2023. 

Dense Inert Metal Explosives (DIME) are typically used in munitions intended to reduce civilian harm. Non-reactive metals, like tungsten, added to the explosives reduce the area impacted by the blast, but increase the power. Often munitions filled with DIME replace steel casing with carbon fibre to reduce fragmentation.

Some sources refer to DIME as a multiphase blast explosive, a term that also covers some explosives with reactive metals. Photos from testing show mannequins near the blast coated in tungsten powder.

Some claims of DIME use in Gaza mention the presence of powder or microscopic shrapnel found on victims. “Peppering” and “tattooing” are mentioned (warning: graphic content) as common injuries in blast victims, where the explosion propels small debris like sand into the body, along with fragments of various sizes.

The US Air Force has accepted delivery of at least 500 DIME-filled GBU-39A/B bombs, and has used at least 23 in combat. No transfers of GBU-39 A/B FLM bombs from the US to any other country, including Israel, have been reported, and a Bellingcat analysis of GBU-39 strikes in Gaza between October 2023 and January 2026 did not find any evidence of this variant being used.

There is currently no conclusive evidence that militaries aside from the US have used DIME in combat.

Clues From Clouds

Clouds, and the colours of the smoke can provide clues about the type of explosive. However, chemical composition, environmental conditions, and location can all affect how explosions appear. 

Clouds

This footage, originally posted on social media in November 2025, shows an explosion in Gaza.

The Israeli army launched thermobaric and pressure bombs, supplied by the United States, on Gaza. These bombs, which burn at a temperature of 3,500 degrees Celsius, are capable of killing thousands in seconds, leaving no trace. pic.twitter.com/pZhoIfsazP

— China pulse (@Eng_china5) February 12, 2026

Video of an explosion in Gaza, falsely attributed as a thermobaric weapon. Source: X/@Eng_china5.

The visible cloud in the video is a condensation or Wilson cloud, caused by an explosive shockwave interacting with humid air. This same effect is visible in videos of the Beirut explosion in 2020, when ammonium nitrate exploded at the port after a fire.

Another view of the explosions in Beirut pic.twitter.com/efT5VlpMkj

— Borzou Daragahi (@borzou) August 4, 2020

Video of the 2020 Beirut ammonium nitrate explosion. Source: X/Borzou Daragahi.

Smoke colours

Colours in the smoke of an explosion can help identify the gases, which in turn can help identify the explosive material, Dr Rachel Lance told Bellingcat. “Yellow, orange, and red tones each indicate the presence of specific chemicals.” 

Black smoke means “the bomb produced a lot of fire and inefficiency, because materials burned instead of detonated, and was probably a homemade or improvised explosive”. White or light grey smoke indicates “an efficient detonation, and that tells us it was a pure, high-grade material inside,” Lance said.

Some munitions, like cruise or ballistic missiles, may have efficient high explosives, as well as low explosive propellants or fuel. The area targeted, such as buildings, may lead to dust or debris that obscure the gases created by the explosion. 

In some cases, multiple bright fireballs are launched into the sky, accompanied by a rapid humming or throbbing sound and bright flashes. This typically happens when solid-fuel rocket motors, like those in air defence or ballistic missiles, are burning or exploding.

Major secondary explosions after a U.S. airstrike in the vicinity of Higuerote Airport in Venezuela tonight. pic.twitter.com/NrFOVj9IfM

— OSINTtechnical (@Osinttechnical) January 3, 2026

Venezuelan Buk Air Defense System rocket motors ‘cooking off’ after being targeted by US strikes in Jan. 2026. Source: X/Osinttechnical.

Geolocation

Geolocation of the explosion site can help identify or rule out potential explanations. Large explosions can be caused by much smaller bombs hitting storage sites or production sites for ammo. The geolocation of the video below indicated that the location hit was a storage area for missiles.

Qom today looks like it was hit by a GBU 57 bunker buster.

The GBU 57 Massive Ordnance Penetrator is a 30,000 pound bunker busting bomb designed to penetrate deep underground before detonating. pic.twitter.com/d4bGJ19nQb

— Open Source Intel (@Osint613) March 11, 2026

Video shared by a user claiming this video shows the use of a GBU-57 “Massive Ordnance Penetrator”. A now-suspended user claimed the video showed the “Mother of All Bombs”. Source: Osint613.

Blast Effects on People

Misinformation regarding blast effects on people might lead to reports of harm to be wrongly dismissed or false claims about mystery weapons to spread.

In February 2026, claims of “vaporisation” or disintegration of people due to thermobaric weapon explosions appeared online. Days later, counterclaims argued that explosives can’t “disintegrate” people and thermal effects were not responsible.

According to multiple studies, even less powerful explosives can cause disintegration. When explosions occur in enclosed spaces, such as inside a building, they reflect shock waves, leading to increased blast effects.

Blast injuries are generally classified into four categories, based on what mechanism is causing the injuries.

The primary effect, the blast itself, “puts tremendous strains on human tissue, causing them to rip and tear, both internally and externally, so massive internal bleeding can occur,” Brian Castner, a weapons investigator for Amnesty International, told Bellingcat.

Primary injuries can lead to a variety of symptoms, including vertigo, vomiting blood, and bleeding from the ears. A viral post shared by the White House Press Secretary claimed to be firsthand testimony from a Venezuelan security guard following US strikes in Venezuela. The post alleged that the US used a sonic weapon without any supporting evidence, and the symptoms described are typical of primary blast injuries.

The secondary effect results from the metal fragments of the munition. Some weapons are specifically designed to break into uniform small pieces, Castner said. “Even small fragments, the size of a bullet, can break a bone, since the metal is flying through the air so quickly,” the weapons investigator explained.

Even single fragments can injure or kill people hundreds of metres away from a blast. People close to it may be largely disintegrated, often described (warning: graphic content) as “total body disruption” in Forensic Medicine.

“Combined, these blast and fragmentary effects can do horrific damage to the human body, and if a person is close enough to a large munitions detonation, leave little trace they ever existed,” Castner told Bellingcat.

A recent Bellingcat investigation into three specific US-made munitions used in Gaza found videos showing small pieces of human bodies consistent with total body disruption, at several different strikes within the dataset.

Explosions can also cause burns or thermal injuries. Temperature is not the most relevant factor, because “by the time a human body is exposed to the temperatures of a burning explosive, people will have severe trauma and death,” Dr Lance told Bellingcat.

In many real-world cases “the blast pressure reaches farther than the thermal flash,” Dr Sabrina Wahler said. “The thermal danger becomes much larger and longer lasting when the explosion occurs in a confined space, when the formulation supports continued burning with air, or when the detonation triggers secondary fires that keep generating heat well after the initial blast,” she noted.

Flash burns are often seen on exposed parts of the body close to the blast (warning: graphic content). Explosions that start fires or contain incendiary materials can result in severe burns.

Are These Explosives Legal?

Misinformation often raises questions about legality, with false claims that specific weapons are inherently illegal or misrepresenting how they work. This is one of the reasons that nations conduct legal reviews of new weapons, Michael Meier, a former Senior Advisor to the Army Judge Advocate General for Law of War, and current Adjunct Professor at Georgetown University Law Center, told Bellingcat.

Thermobarics and DIME are legal if their use complies with specific principles of international humanitarian law (IHL) and the law of armed conflict (LOAC), such as proportionate and discriminate use, experts told Bellingcat.

“Even lawful weapons can be used in an unlawful manner”, Michael Meier said. One example is when they are directed against civilians or when they are used in a manner that breaches the principles of distinction or proportionality, he explained.

“The law’s ability to prevent harm is constrained by the compromises between military necessity and humanity made in its creation,” Dr Arthur van Coller, Professor of International Humanitarian Law at the STADIO Higher Education and a legal expert on thermobaric explosives, told Bellingcat.

“As a result, weapons that cause immense destruction may remain lawful (even nuclear weapons) if they fit within legal definitions, even when their humanitarian impact is severe,” van Coller explained.

Bellingcat is a non-profit and the ability to carry out our work is dependent on the kind support of individual donors. If you would like to support our work, you can do so here. You can also subscribe to our Patreon channel here. Subscribe to our Newsletter and follow us on Bluesky here, Instagram here, Reddit here and YouTube here.

The post Explosive Misinformation: A Guide to Mushroom Clouds, ‘Sonic Weapons’ and Disintegration appeared first on bellingcat.

Flight tracking data is an important tool in open source research, but with 100,000 daily flights, it can be difficult to contextualise what a particular aircraft’s movements indicate. 

Bellingcat has developed a tool called Turnstone to make it easier to visualise historical trends in flight data and spot unusual patterns. It also allows users to filter by parameters such as aircraft type or a geographic region of interest. 

This tool primarily uses Automatic Dependent Surveillance–Broadcast (ADS-B) data, the technology that enables open source investigators and enthusiasts to track flights. 

Most aircraft are equipped with transmitters that broadcast ADS-B data to comply with global aviation regulations, though regulations vary by jurisdiction, and military aircraft might not always transmit. ADS-B data includes information about an aircraft’s identity and type, as well as its precise position, speed and altitude. 

Popular flight-tracking websites such as Flightradar24 and ADS-B Exchange typically display historical data for a particular time or aircraft. However, Turnstone aggregates ADS-B data for multiple aircraft over time, and allows users to search for flights across two areas of interest at once. These features provide additional context for open source investigators to better understand flight behaviour.

Watch the video for a demonstration of how the tool works, using the example of Black Hawk helicopter patrols near one of the borders between the US and Canada:

You can view Turnstone’s source code and information about hosting it yourself on Bellingcat’s GitHub. 

We also have a web-based instance of the tool that journalists and academics can access. Due to data hosting and processing costs, we can only grant access on a selective basis. If you would like to apply, please fill in this form. Priority will be given to researchers conducting open source investigations aligned with Bellingcat’s goals.

Read on for more examples of how Turnstone can be used for investigations, as well as some limitations of the tool.  

Spotting Unusually High US Tanker Activity Before Iran Strikes

The US and Israel launched joint air strikes across Iran on Feb. 28, 2026, reportedly killing more than 1,000 people, including members of the Iranian leadership, in five days.

This marked a dramatic escalation since the US and Israel bombed three Iranian nuclear sites in June 2025. 

Flight data before both the June 2025 and February 2026 strikes showed a large number of American aerial tankers leaving the US and crossing the Atlantic towards Iran. Aerial tankers such as the KC-135 and KC-46A can refuel military aircraft in-flight, making them essential for most long-range combat missions.

With Turnstone, it is possible to interrogate the baseline level of movement and see how unusual this activity is.

To do this, three filters are set on the search: a geographic region of interest, set to the North Atlantic, a filter on the aircraft type, to search only for tankers, and a filter on the aircraft heading, to search only for eastbound traffic.

[Note: For the aircraft category designations, Bellingcat used a custom-prompted large language model (LLM), Claude Sonnet 4.0, to assign a category label using aircraft type code data. There may be some inaccuracies in the classifications, as LLMs are prone to hallucinations. We discuss this further in the “Limitations of the Data” section of this piece.]

This search finds over 40,000 aircraft locations that match these filter queries. However, a look at the summary table shows that this data includes non-American tankers as well.

We can filter this data to include only aircraft associated with the US by typing “United States” into the search box in the table. Note that ownership data is not 100 percent accurate – it may be out of date, especially for privately owned aircraft, and new aircraft might not have any data at all. However, especially when comparing trends over time or searching for research leads, this data can still be useful.

The graph of matching detections over time now shows that while there is a large baseline level of transatlantic movement for American tankers, there was a notably higher number of American tankers heading eastward from the US across the North Atlantic detected in the week of June 15, 2025, as well as in the last two weeks of February 2026.

A week after the increased eastbound traffic in June 2025, early in the morning on June 22, the US struck several nuclear sites in Iran. And on Feb. 28, 2026, the US and Israel launched over 900 strikes against Iran.

Altering the search query to look for westbound tankers instead of eastbound tankers, we can also see a larger-than-normal number of American tankers heading in the direction of the US during the week of July 13, 2025, bookending the summer airstrikes in Iran. No such return movement is yet visible following the recent strikes.

Finding Deportation Flights to Guantanamo Bay

Turnstone also allows you to search for aircraft detected across two different geographic regions of interest (ROIs). 

Shortly after US President Donald Trump announced the opening of a migrant detention centre at Guantanamo Bay in Cuba at the end of January 2025, the US military reportedly flew about 100 immigrants from El Paso, Texas, to the US naval base to await deportation. By selecting the areas around both Guantanamo Bay and El Paso, we can find flights between these cities that broadcast ADS-B data.

When two ROIs are selected, you can also enter the maximum time difference between an aircraft’s presence in the two regions. 

In the example below, we have entered 36,000 seconds (10 hours), meaning that the aircraft must have crossed through both regions within 10 hours of each other. We have also set the maximum altitude to 15,000 ft (4.57km) to look for planes landing and taking off. This limit is set relatively high as there are no ADS-B receivers at Guantanamo Bay, and only the initial approach is captured.

After five months with no tracked flights between the two locations, this search shows an uptick in flights in the few months from February 2025.

Results for this search query from Jan. 26, 2026, include several passenger aircraft operated by companies known to run deportation flights from the US, such as Omni Air International and Global Crossing Airlines.

Mapping US Customs and Border Patrol Aircraft

Turnstone also supports uploading a list of International Civil Aviation Organization (ICAO) addresses, informally referred to as aircraft “hex codes”, which are unique identifiers assigned to aircraft by ICAO member states.

For example, to explore data related to Department of Homeland Security (DHS) activity and look for patterns related to the US immigration enforcement and border security operations, we can copy and paste the hex codes from a list of US Customs and Border Patrol (CBP) aircraft (used across the DHS) into a text file, and upload that file. Now, we can search among these aircraft with any of the same filters demonstrated in the earlier case studies. Alternatively, we can also deselect all of the filters to track the most recent activity by those aircraft.

Let’s try that with the CBP list, this time with a very large number of results selected: 500,000. Note that increasing the number of results increases the search time and requires more browser memory.

When many points are displayed, the map is simplified, and hover features are disabled.

By the California-Mexico border, Eurocopter AS350 (type “AS50”) can be seen on frequent patrol missions over the land border. Over the Pacific Ocean, Black Hawk helicopters (“H60”) can be seen patrolling the international waters boundary off the Mexican coast, while CBP Dash-8s (“DH8B” and “DH8C”) travel farther offshore.

In contrast, by the Minnesota-Canada border, CBP makes more active use of one of its MQ-9 Reaper drones, as seen from the prevalence of red dots that correspond to “Q9”, the type code of these drones, in the results map.

Let’s take a closer look at these drones by filtering the results with the text “Q9”. Now the displayed aircraft only include MQ-9 Reaper drones.

Now we can take a closer look at the patterns of drones, specifically among the search results.

While overall CBP flight activity was relatively stable, drone flights seem to have intensified in December 2025 and January 2026, compared with previous weeks.

Limitations of the Data

In open source research, it is always important to be alert to the limitations of a particular data source, and ADS-B data is no exception. 

For example, some aircraft do not have ADS-B transponders and use older transponders to transmit flight information, which can result in tracking tools such as Turnstone showing inaccurate position data. 

In the previous case study of CBP aircraft, the Turnstone results appeared to show an MQ-9 Reaper drone in Canada on Jan. 20, 2026. 

Is this evidence of covert DHS missions in Canadian airspace? Likely not: a cross-check of the drone’s hex code on that date with ADS-B Exchange shows that the aircraft’s position track is not smooth, but jumps back and forth between a line in the US and several points many kilometres away in Canada.

This happens because when ADS-B position data is not available, flight trackers often use multilateration (MLAT), which estimates the location of the aircraft using the time differences between signals transmitted from known sites, as a substitute. The flight tracking information on ADS-B Exchange shows that the position was calculated using MLAT, which is less accurate than position data directly transmitted through ADS-B. ADSB.lol, which is the data source used by Turnstone, uses MLAT when ADS-B position data is not available.  

ADS-B data is also limited by where ground antennas are available to receive radio signals from aircraft and by when aircraft choose to transmit the data.

Other datasets which Bellingcat has used to enable the filters available on Turnstone each have their own limitations. 

There is no single source of data on aircraft ownership. ADS-B data identifies an aircraft only using its ICAO address or hex codes, but does not contain other information that directly specifies the type of aircraft or its registration.

Instead, flight-tracking websites reference aircraft registration databases, such as those maintained by the US Federal Aviation Administration, to correlate ICAO addresses with registration information. The ownership data displayed on Turnstone is from tar1090-db, a community-maintained project which has produced the most comprehensive freely available global aircraft registration database. However, since ownership data is collected from many jurisdictions, with different privacy and disclosure requirements, it may sometimes be out-of-date or misleading. 

Ownership information displayed in Turnstone or any other flight-tracking software should still be verified independently using multiple sources.

For example, one of the aircraft that came up in the search for flights between El Paso and Guantanamo Bay had a hex code of a6b0f5. This showed up in Turnstone’s results as being owned by Bank of Utah Trustee, which matches the operator listed for this flight on ADS-B Exchange. But some of the flight codes used by this aircraft, starting with “GXA”, are used by Global Crossing Airlines (GlobalX). The Bank of Utah is known to legally own aircraft under a trust relationship, while leasing the aircraft and operational control to third parties such as GlobalX.

The “Category” label and “Military” flag, which provide a convenient way to filter aircraft, are pre-generated by a custom-prompted large language model, Claude Sonnet 4.0, based on the make and model of an aircraft. 

For example, the LLM may take a type code of A321, which refers to an Airbus A321 passenger jet, as input and assign the corresponding aircraft the category of “airliner”. 

Bellingcat manually verified over 80 per cent of aircraft, corresponding to the most common aircraft types. But as we know, LLMs are prone to hallucinations, and categorisation may be inaccurate for more obscure aircraft. Additionally, some aircraft, such as the V-22 Osprey, fall between categories and are inherently ambiguous. 

To prevent errors caused by the potential miscategorisation of aircraft, you may want to search by type code, which will draw from the raw tar1090-db data, rather than category. All aircraft registration, type, and owner information should be independently verified.

Suggestions and Further Information

As we’ve seen in this guide, Turnstone searches historical ADS-B data to allow researchers to explore flight patterns over time and in specific locations. While flight-tracking data has inherent limitations, Turnstone can provide useful leads for researchers looking to incorporate flight tracking in their investigations.

If you have suggestions for improving the tool, you can submit a pull request on Bellingcat’s GitHub. More technical information can also be found in the tool’s README.

For more demos and information about the history of this tool, watch a talk that Bellingcat gave about it at the What Hackers Yearn (WHY) 2025 hacker camp:

Bellingcat is a non-profit and the ability to carry out our work is dependent on the kind support of individual donors. If you would like to support our work, you can do so here. You can also subscribe to our Patreon channel here. Subscribe to our Newsletter and follow us on Bluesky here and Mastodon here.

The post Using Bellingcat’s New Open Source Tool to Explore Historical and Spatial Flight Data appeared first on bellingcat.

To stay up to date on our latest investigations, join Bellingcat’s WhatsApp channel here.

Federal agents have frequently used so-called “less-lethal” weapons against protesters, including impact projectiles, tear gas and pepper spray, since the Trump administration’s nationwide immigration raids began last year. 

The use of less-lethal weapons (LLWs) has been controversial. While designed to incapacitate or control a person without causing death or permanent injury, they can cause serious or fatal injuries, especially when used improperly. 

Earlier this month, two protesters in California were reportedly blinded after US federal agents fired less-lethal rounds at their faces from close range. These incidents were part of a wave of violent clashes between agents from the Department of Homeland Security (DHS) and protesters across the country after the deadly shooting of US citizen Renee Good by an Immigration and Customs Enforcement (ICE) agent in Minneapolis. 

In protests in Minneapolis immediately following Good’s death, one Customs and Border Patrol (CBP) officer was captured on camera firing a 40mm less-lethal launcher five times in less than five minutes, with several of these shots appearing to target protesters’ faces, which is against CBP’s own use-of-force policy. 

A Bellingcat investigation of DHS incidents in October 2025 also found about 30 incidents that appeared to violate a temporary restraining order (TRO) issued by an Illinois judge restricting how DHS agents could use LLWs.

Support Bellingcat

Your donations directly contribute to our ability to publish groundbreaking investigations and uncover wrongdoing around the world.

Donate

It is not always obvious whether the use of a LLW is authorised or not, as DHS component agencies such as ICE and CBP have varying guidance on factors such as the level of resistance an individual needs to show before a certain type of force can be used, as well as how specific types of less-lethal weapons and munitions can be used. 

While CBP’s use-of-force policy as of January 2021 is available on its website, ICE does not include specific guidance on less-lethal weapons in its 2023 “Firearms and Use of Force” Directive, and does not appear to have any publicly available policy that outlines this guidance.

DHS did not respond by publication time to Bellingcat’s request for the most recent DHS, CBP and ICE use-of-force policies, or to questions about what less-lethal weapons were authorised for use by the department and its component agencies. 

The DHS use-of-force policy, updated in February 2023, states that the department’s law enforcement officers and agents may use force, including LLWs, “only when no reasonably effective, safe and feasible alternative appears to exist”. It also says agents may only use a level of force that is “objectively reasonable in light of the facts and circumstances” that they face at the time.

DHS has repeatedly defended its use of riot-control weapons in protests across the country, stating that it was “taking reasonable and constitutional measures to uphold the rule of law and protect [its] officers”. 

Here’s how to identify some of the less-lethal weapons that DHS agents, including those from ICE and CBP, have been seen using during recent immigration operations. 

Compressed Air Launchers or ‘PepperBall Guns’

Compressed air, or pneumatic launchers, are essentially paintball guns that fire 0.68mm balls which break on impact. Often, this releases a powdered chemical irritant such as oleoresin capsicum (OC) or PavaPowder – the same compounds typically found in pepper spray. 

Compressed air launchers can also be used with other projectiles, such as “marking” projectiles that use paint to mark an individual for later arrest, and projectiles intended to break glass.

These weapons are often referred to as “PepperBall” guns, named after the leading brand PepperBall. However, DHS agents have also been seen carrying compressed air launchers from different brands, such as the FN303, produced by FN America.

Many compressed air launchers resemble standard paintball guns, with a distinct hopper or loader, which holds the ball projectiles, mounted to the top. They also have a compressed air tank that might be mounted to the side, bottom, or inside the buttstock (or back) of the weapon.

Many compressed air launchers, and less-lethal weapons in general, have very bright colours such as orange to distinguish them from lethal weapons. 

However, some compressed air launchers require closer scrutiny to distinguish them from firearms. 

For example, federal agents have been seen carrying FN303 compressed air launchers in videos of immigration enforcement activities. This weapon may resemble a rifle or other firearm, as it is usually all-black and, unlike the TAC-SF series PepperBall guns, lacks a visible hopper. 

If closer examination is possible, this weapon can be identified by its distinct features, including a circular magazine, side-mounted compressed air tank and a hose connecting the firearm to the air tank.

The January 2021 CBP Use of Force Policy places several restrictions on the use of compressed air launchers, including that they should not be used against small children, the elderly, visibly pregnant women, or people operating a vehicle. It also states that PepperBall guns should not be used within 3 feet “unless the use of deadly force is reasonable and necessary”. When using the FN303, the minimum distance is increased to 10 feet. 

The CBP Use of Force Policy says that the intentional targeting of areas where there is a “substantial risk of serious bodily injury or death is considered a use of deadly force.” Agents are instructed not to target “the head, neck, spine, or groin of the intended subject, unless the use of deadly force is reasonable”. PepperBall and FN America provide similar warnings about avoiding vital areas to prevent serious injury or death.

According to a 2021 report by the US Office of Inspector General, CBP requires its agents to recertify their training to use PepperBall guns and FN303s every year, but ICE does not.   

40mm Launchers

DHS agents also use 40mm launchers to fire “Less-Lethal Specialist Impact and Chemical Munitions (LLSI-CM)”. These launchers resemble military grenade launchers, but are used to fire less-lethal ammunition, including “sponge” rounds that can disperse chemical irritants on impact. 

Federal agents have been seen using or carrying the B&T GL06 launcher in footage of multiple incidents reviewed by Bellingcat. They have also been spotted with other 40mm launchers, including Penn Arms 40mm multi-shot launchers, which have a six-round cylinder magazine. 

There are various less-lethal munitions available for 40mm launchers, including those whose primary function is “pain compliance” through the force of impact, chemical irritants or a combination of both. 

Videos of clashes between Border Patrol agents and protesters show these launchers being used with combination rounds designed to hit the target for pain compliance while also delivering a chemical irritant such as OC or CS. 

Other munitions dispense chemical irritants or smoke after being launched. For example, in the protests immediately following Good’s death, a Border Patrol agent was seen firing a 40mm munition that released multiple projectiles emitting chemical irritants in a single shot, consistent with the “SKAT Shell” by Defense Technology.

Defense Technology’s technical specifications for its 40mm Direct Impact Rounds, which agents have been seen armed with, state that the munitions are considered less-lethal when fired at a minimum safe range of 5 feet and at the large muscle groups of the buttocks, thigh and knees, which “provide sufficient pain stimulus, while greatly reducing serious or life-threatening injuries”.

A DHS Office of Inspector General Report in 2021 noted varying guidance on the use of 40mm launchers among the department’s component agencies: “ICE’s use of force policy indicates that the 40MM launcher is deadly force when fired at someone, while the CBP use of force policy only directs officers not to target a person’s head or neck.”

CBP’s 2021 use-of-force policy states that agents should “not intentionally target the head, neck, groin, spine, or female breast”, and that anyone in custody who has been subject to such munitions should be seen by a medical professional “as soon as practicable”.

As of publication, DHS had not replied to Bellingcat’s questions about whether the department had an internal policy or provided training to staff on the minimum safe distance for 40mm less-lethal launchers as recommended by the manufacturers.

Hand-Thrown Munitions

DHS agents have also been seen throwing some less-lethal munitions, such as flash-bangs, smoke and “tear gas” grenades or canisters by hand. 

Related articles by Bellingcat

Investigations

‘Pattern of Extreme Brutality’: Tear Gas, Pepper Balls Among Weapons Deployed Against Protesters in Illinois

These munitions activate a short delay after the grenade is employed. When they activate, flash-bangs or “stun” grenades emit a bright flash of light and a loud sound that is designed to disorient targets. Both smoke grenades and tear gas (also known as “CS gas” or “OC gas”) emit thick smoke, but the former just impedes visibility, whereas the latter also contains chemical irritants that sting the eyes. 

Defense Technology offers smoke grenades with hexachloroethane smoke composition, but most of their smoke grenades use “SAF-Smoke”, a less toxic terephthalic acid smoke composition. 

Hexachloroethane, while toxic, is not a nerve agent, despite misinformation surrounding the deployment of green colored smoke grenades in Minnesota by DHS personnel. 

The shape and general construction, colour, and any text can help identify these munitions.

Less-lethal munitions typically feature the manufacturer’s logo, the model name of the munition, and the model or part number. The text and manufacturer logo are typically colour-coded to indicate the type of payload the munition has, with blue indicating CS, orange indicating OC, yellow indicating smoke, green indicating a marking composition and black indicating munitions with no chemical payload. 

A 2021 analysis by Bellingcat and Newsy found that Defense Technology and Combined Tactical Systems, the two manufacturers which produce most of the less-lethal munitions used by federal agents, both list the model numbers of their products online. Publicly available price lists for Defense Technology and Combined Tactical Systems can also be used to identify specific munitions by their model numbers. 

CBP’s 2021 use-of-force policy states that hand-thrown munitions are subject to the same restrictions for use as munition launcher-fired impact and chemical munitions. 

Chemical Irritant Sprays

DHS agents have also been using handheld chemical irritant sprays, often colloquially referred to as “pepper spray” or “mace”.

These sprays come in a variety of sizes and concentrations containing CS, OC, or both. Sprays used by law enforcement usually have a canister size designated “MK-” followed by a number, with higher numbers indicating larger canister sizes. The concentration of chemical irritants contained in the spray is also indicated on the canister.

The effectiveness of OC sprays is determined by the concentration of major capsaicinoids, which are the active compounds in OC that cause irritation. These sprays are also affected by the type of aerosol dispersion, or stream, used. Different types of streams increase or decrease the range of the spray as well as the coverage area. 

Civilian and law enforcement sprays range from 0.18 percent to 1.33 percent major capsaicinoids, according to SABRE, a producer of law enforcement and civilian sprays. Civilian sprays in the US can have the same major capsaicinoid content as law enforcement sprays, but are restricted to smaller-sized canisters. 

Defense Technology sprays have different colour bands to indicate the percentage of major capsaicinoids in the spray for OC. If the spray is CS, the CS concentration is standardised at 2 percent. The company uses a white band for .2 percent, yellow band for .4 percent, orange band for .7 percent, red band for 1.3 percent and a grey band for sprays containing either CS or a combination of OC and CS.

SABRE sells a variety of concentrations and sprays as law enforcement products, including 0.33 percent, 0.67 percent, and 1.33 percent major capsaicinoid concentrations of OC, as well as CS, and combination CS and OC sprays. The specific concentrations of SABRE sprays and the type of stream can also be identified by the text on the canister. 

One Air Force Research Laboratory study found that some sprays may pose a significant risk of severe eye damage due to pressure injuries resulting from large aerosol droplets hitting the eye. 

Defense Technology’s technical specifications recommend a minimum distance of between 3 and 6 feet, depending on the specific spray. SABRE does not publicly provide their minimum safe deployment distances, but a Mesa Police Department document lists a minimum distance of six feet for the SABRE Red MK-9. CBP’s 2021 use-of-force policy does not provide any minimum use distances. 

CBP’s 2021 use-of-force policy states that OC Spray may only be used on individuals offering “active resistance”, and that it should not be used on “small children; visibly pregnant; and operators of motor vehicles”. 

Electronic Control Weapons

DHS agents have also been seen using electronic control weapons (ECWs), which are colloquially called TASERs after the original weapon invented for law enforcement use, in immigration-related raids. 

ECWs can deliver a shock upon direct contact or launch probes that embed in the targeted person, incapacitating them. 

A shock on contact, or a “drive-stun” feature, delivers localised pain while in direct contact. When properly deployed, the probes send signals to the body that cause muscles to contract. A person’s body “locking up” from muscle contractions is an indicator that an ECW has been deployed. ECWs may be capable of using either or both methods.

ECWs are typically painted a combination of black and bright yellow, but this varies between models. The bright colour of parts of tasers is a common feature to help distinguish an ECW from handguns used by federal agents. When viewed from the front, a circular gun barrel is visible on handguns, while ECWs feature multiple circular probes or rectangular covers on the cartridge. ECWs also usually have flashlights and lasers, although handguns may also be equipped with these features. Some ECWs may make audible sounds when armed or deployed.

Axon, the predominant manufacturer of ECWs, produces several models including the TASER 10 and TASER 7. Axon provides a policy guide on recommended use of its TASER models to law enforcement agencies, which recommends targeting below the neck from behind, or the lower torso from the front. It recommends avoiding sensitive areas including the head, face, throat, chest and groin. 

Axon also recommends against using ECWs against small children, the elderly, pregnant people, very thin people and individuals in positions of increased risks such as running, operating a motor vehicle, or in an elevated position “unless the situation justifies an increased risk”.

CBP’s 2021 use-of-force policy, in addition to restricting the use of ECWs against small children, the elderly, visibly pregnant women, and people operating a vehicle, states that they should not be used against someone who is running or handcuffed. However, the policy does state that there may be an exception to the rule against using ECWs on a running person if an agent has a “reasonable belief that the subject presents an imminent threat of injury” to an agent or another person. This threat, according to the policy, must “outweigh the risk of injury to the subject that might occur as a result of an uncontrolled fall while the subject is running”.

Bellingcat is a non-profit and the ability to carry out our work is dependent on the kind support of individual donors. If you would like to support our work, you can do so here. You can also subscribe to our Patreon channel here. Subscribe to our Newsletter and follow us on Bluesky here and Mastodon here.

The post Identifying ‘Less-Lethal’ Weapons Used By DHS Agents in US Immigration Raids and Protests appeared first on bellingcat.

Join Bellingcat’s WhatsApp Channel for the latest news and resources from us.

How does one monitor a conflict zone on the brink of civil war, especially in a region which is difficult to access, experiences frequent internet shutdowns and where misinformation is common? In this guide, we outline the open source tools and methods we can use to evidence what is really happening in many such conflict settings.

Our focus for this guide is on India, which recorded 84 internet shutdowns in 2024 – the highest number amongst democratic nations. In early June, authorities imposed a curfew and suspended internet access in parts of Manipur after protests erupted over the arrest of ethnic leaders. The state, in the north-east of the country, has been wracked by violence for years.

The ethnic conflict between the majority, mostly Hindu Meitei population and the indigenous, largely Christian Kuki Zo communities is one of the worst spates of violence Manipur, also known as the “Land of Jewels”, has experienced in decades. 

The Imphal valley in Manipur is surrounded by mountains. It is home to 39 ethnic communities. Just over half of its nearly three million residents belong to the Meitei community, followed by the Naga (20 percent) and the Kuki Zo (16 percent) tribes.

The landscape is complex, with ethnic armed groups divided into multiple factions (this list is not complete):

• The ethno-nationalist militia – yet to be designated as a banned group – Arambai Tenggol (AT), the United National Liberation Front (UNLF) – Meitei
• Kuki National Army, Kuki National Front – Kuki 
• Zomi Revolutionary Army – Zomi
• National Socialist Council of Nagalim (Isak Muivah) – Naga

In May 2023, the Manipur High Court passed an order recommending a Scheduled Tribe status (a category for indigenous communities in India that guarantees affirmative action and constitutional protection over identity and land) for the dominant Meitei community. Tribal communities rallied against the decision while the Meitei community held counter-rallies and counter-blockades. Clashes broke out between the Kuki and Meitei groups. Since then, the conflict has displaced more than 60,000 people and claimed more than 260 lives from both communities.

In this guide, we show you how to use open source methods in any secluded area to:

• Analyse weapon imagery and the groups using them
• Investigate weapons that were looted and where they ended up
• Analyse images of drones potentially used as weapons to deploy munitions

Analysing Weapon Imagery

One effective approach for open source researchers is to trace the digital footprint of  weapons. In the Manipur case, local armed groups, such as the Arambai Tenggol, the UNLF and the Kuki National Front, have been posting weapon imagery mainly in WhatsApp groups and Facebook accounts.

Support Bellingcat

Your donations directly contribute to our ability to publish groundbreaking investigations and uncover wrongdoing around the world.

Donate

According to media reports, the war has been fueled by weaponry looted from police armouries or procured on the black market either from Myanmar, across the border, or through surrenders in amnesty drives.

The 6,000 firearms looted included pump-action shotguns, grenade launchers, AK-pattern rifles, INSAS rifles and ammunition. The police claimed that in February and March alone, more than 1,000 weapons were surrendered, with more than half from the Meitei-dominated valley districts, where a majority of the weapons were looted.

Bellingcat analysed weapon imagery from 2023 and 2024, accessed from WhatsApp groups and Facebook accounts linked to non-state actors, including the AT, Kuki Zo militant groups, and various volunteer organisations. While these groups have surrendered some weapons in amnesty drives, many sophisticated weapons were not turned in and were only recovered in search operations by security forces.

For verification, we ran screenshots of images of the weapons without visible serial numbers or other markings first through reverse image searches on Google and Yandex. Then, we cross-referenced the images with resources like the Small Arms Survey handbook and Open Source Munitions Portal (OSMP). 

However, these databases are limited in their documentation from India.

We also looked at the public dashboard of Conflict Armament Research (iTrace). This is a far larger data source. However, the full dataset, which contains a huge number of images of weapons from around the world, is not publicly available. Only broad statistics, and no images, are visible via the dashboard.

The Small Arms Survey handbook helped match and identify, to an approximate accuracy, some of the older weapon models published on social media platforms and YouTube. However, the guerrilla modifications or customisation of weapons by the militant and militia groups made it challenging to identify the specific models.

This was the case in a video posted on X, which purported to show militants preparing to fire a mortar projectile.

By breaking down the video into frames using InVID, a platform that contains a number of useful tools for analysing videos, we were able to identify the weapons, providing clearer imagery we could use to go back to reverse image tools on Google and Yandex, as well as the Small Arms Survey handbook.

We identified three weapons from the video:

• A bolt action rifle
• A FAL (Fusil automatique leger)
• A 60mm commando mortar

The shape of the weapon held by the militant wearing the beanie cap and scarf in the same video matches a FAL pattern rifle, such as the Indian 1A1 FAL, which has a distinctive long wooden handguard with multiple elongated ventilation holes. 

“In India, the rifle was produced by the Ordnance Factory, Tiruchirappalli and was in service up to 1998, when it was replaced by the INSAS Rifle. Over a million units of the 7.62 mm SLR rifle have been produced by the OFB,” wrote (Retd) Major General Dhruv C Katoch, who previously served as the Director for Centre for Land Warfare Studies.

Also visible in the video is an unidentified model 60mm commando mortar. Commando mortars are characterised by a more portable design, typically featuring a much smaller baseplate and a sling or carrying handle rather than a bipod, all of which can be seen in the images below.  The reverse image search on Google led us to a file photo on Wikimedia posted by the US Army, besides this assessment by Jesus Roman, Editor of Revista Ejercitos.

Munitions researcher and PhD candidate in War Studies at Kings College London, Andro Mathewson, described it as likely being a 60mm mortar. “It looks like one man is using the mortar tube, which is relatively unusual. Normally, it’s at least a two or three-man team. And the munition looks light green in colour with a sort of light metal-coloured fuse and light silvery tail fins,” he said. “It’s definitely a small calibre mortar, which is a mainstay in military forces. This appears to be military/official manufacture rather than improvised,” Matthewson told Bellingcat.

Which Groups Use the Weapons?

From the data collected from 2023 and 2024, Bellingcat found that many rifles in the images have different furniture and display cloth wraps, improvised slings, aftermarket optics, even taped-on foregrips.

The next step is to identify the various groups in the pictures. Analysing symbols is a good way to do that. For example, we know that the Saipikhup is the traditional weave of the Kukis. It symbolises heritage and identity and is often worn during important occasions. We also found images of Kuki militants wearing this handwoven shawl (saipikhup) belonging to the Thadou indigenous tribe.

Their fatigues bear the insignia of the President faction of the Kuki National Front, which has been accused of attacking paramilitary security forces. Meanwhile, the same group in the image brandishes AR-pattern and INSAS rifles. The INSAS rifle is an Indian police or military issue, matching reported looting from armouries. Several weapons in the image were also heavily customised, consistent with militia or irregular combatant practices.

Other images also offer clues.

The Kangleipak is a seven-colour flag usually brandished by the AT.

Bellingcat also identified the AT’s commander-in-chief, Korounganba Khuman, in the photos and videos. He actively posts on his Facebook profile and has been widely covered by the local and national press.

News outlets are also valuable sources of information. They might contain images of symbols such as flags, which you can then search for on social media. In one of the videos, we identified militants speaking Meitei Lon, a language used by Arambai Tenggol and militant groups like the UNLF, preparing to fire a mortar projectile from a mortar. Their fatigues bore an insignia that we matched to the UNLF armed group using reverse image search, which led us to a news story featuring the group’s flag as the lead image.

Praveen Donthi, a senior analyst with the Crisis Group who visited Manipur last year during elections, told Bellingcat that though he hadn’t seen arms on any of the aforementioned Kuki Zo militant groups, he had seen INSAS rifles, automatics and double-barrel shotguns being wielded by several young men in the Imphal valley.

“I saw these young men who must have been in their early teens to early twenties when I’d gone to meet Meitei Leepun [a Hindu right-wing activist group] first wielding what looked like state-issued weapons,” he said. “Then later, they replaced it with double-barreled guns. But their leader [Pramot Singh] was openly carrying a pistol in his holster when he came to meet me,” Donthi explained.

Donthi, a former journalist who has reported from conflict zones in Kashmir and Chhattisgarh in India, said he was struck by the young men who were heavily armed in a volatile environment without any evident goal or political ideology guiding them.

Weapons Looted: Where Did They End Up?

When investigating conflicts, identifying the origin of weapons is one of the most difficult tasks, particularly in regions plagued by misinformation or a lack of reliable data. This is the case in Manipur.

Of the 6,000 firearms and ammunition looted from state police armouries mentioned earlier, about half of the weapons have been recovered to date. Around 1,200 matched serial numbers from official inventories, according to reports. Of the weapons recovered, approximately 800 sophisticated ones likely originated outside the state, and 600 were crude, locally produced firearms.

The largest surrender of weapons took place in February and March when more than 1,000 weapons were reportedly surrendered, with more than half from the Meitei-dominated valley districts, where a majority of the weapons were looted. The largest cache was surrendered on February 27 by the Arambai Tenggol (AT). However, the state police is yet to complete categorising the details of the weapons and ammunition surrendered between February 20 and March 6 against the inventory of weapons looted from the state armouries.

Bellingcat requested official data from the Manipur police on surrendered weapons matched against serial numbers from official inventories, but had received no response by the time of publication.

Instead, we decided to see what we could find by using open sources. First, we scraped the state police force’s official X profile (@manipur_police) from Sept. 10, 2023, until June 14, 2025. We did it manually and using Meltwater – a social media monitoring tool.

We dug deeper into media reports, experts’ posts and research to understand what was being used locally. In the Manipur case, recovered firearms include locally manufactured bolt-action rifles, improvised mortars and weapons such as the “Pumpi” – a gun made from repurposed metallic electric poles. These are especially common in the hill areas where the Kuki Zo people live.

The heavy reliance on grenades and improvised explosives is consistent with the guerrilla-style, asymmetric engagements – hit-and-fades, booby-traps, and area denial – rather than large-scale firefights. The presence of multiple improvised munitions types reflects local workshops or village-level bomb-making, likely to supplement limited access to military-grade ordnance, consistent with media reporting on the same (see here and here).

Claims Over Weaponised Drones

In September last year, Indian media reported villagers in the valley district witnessing drones allegedly dropping as many as 50 bombs. Kuki Zo village volunteers and insurgent groups were reported to have set up bunkers in the hills, much like their Meitei counterparts in the valley.

These claims were supported by a Manipur Police statement. The central counter terrorism law enforcement authority, the National Investigating Agency, which filed a case alleging weaponised drone attacks, told the Manipur High Court that Kuki militants dropped 40 drone bombs.

A source in the Defence Ministry told Indian news site The Print that the drone videos circulating online were from either Myanmar or Palestine. Many of the videos showed fertiliser drones, but these were deployed by the People’s Defence Forces in Myanmar, they added.

How Do You Investigate Potential Drone Usage With Open Source Tools?

The Manipur Police posted an image of a drone recovered in the Kangpokpi District, a day after the first set of attacks.

The first step is to identify the possible drone type. The easiest way is by using Google’s reverse image search engine. We identified the drone as commercial-grade, weighing approximately 181g. These carbon fibre lightweight drones, built for speed and agility with a payload capacity of up to 1.5kg, are widely available on the internet. Security sources told The Print that the bombs weighed 300-400g and were nine to 10 inches (23 to 25cm) in size.

After establishing the possible drone type, we can also examine the reported impact sites. Since we only have images of the attack sites shown in the media, we asked Andro Mathewson, a reputed munitions researcher and explosives expert completing his doctoral studies on weaponised drones in smaller conflicts at King’s College London, for help. 

He told Bellingcat that in this situation, “the payload is probably quite small. So the damage won’t be extensive”.

“Some of the images that are shared in The Print report,” added Mathewson, “obviously show a lot of destruction, but a lot of it seems to be sort of secondary destruction from fires rather than from explosions itself”.

Nothing was visible that could specifically determine if drones were used to deploy munitions. The damage from a smaller payload like 400-600g of a grenade would not exceed more than 20 to 30m, according to Mathewson, adding that larger or heavier payloads are not typically seen among non-professional militaries.

The next step is to find out if the munitions have been adapted for drone deployment. Mathewson told Bellingcat that photos of drone parts published by the media were not consistent with munitions deployed by drones. Bellingcat was not able to independently confirm the source or authenticity of these photos.

“That shrapnel looks large, very thick, and very heavy, which is more consistent with larger artillery rounds or even small missiles,” he said. He also noted that the printed fin “looks quite small”.

“Fins made out of plastic are not likely to be attached to a much larger munition that’s produced that type of shrapnel,” he told Bellingcat, saying that “from the scale that we can get in those images, those don’t seem to add up to me”.

For future reference, when we asked Mathewson what to look out for to confirm the use of weaponised drones, he suggested two things. One would be to see and verify videos of drone strikes, either shot by other drones or on phones – something that is conspicuously missing from Manipur despite the authorities’ claims that there have been drone strikes, although there is plenty of online footage of other sophisticated weapons used there. Secondly, Mathewson also said it was worth looking out for 3D-printed munition parts, such as 3D-printed fins that are attached to conventional weapons.

“That’s not necessarily a guarantee, but it’s most closely associated with [modified drones] because the only reason you would attach fins to a grenade, for example, is to make them be dropped from drones,” he added.

Correction: This article has been updated after an image previously incorrectly stated members of the Kuki National Front were posing with AK-47s, M4 Carbines and M16 weapons. We also updated it to reflect errors in identifying the bolt action rifle, the FAL and 60mm commando-style mortar grenade. A section on chronolocation of this video was also removed, and a clarification was added that only the public iTrace dashboard was consulted, rather than cross-referenced with OSMP and the Small Arms Survey.

Additional reporting by Douminlien Haokip.

Pooja Chaudhuri, Claire Press and Gyula Csák contributed to this report for Bellingcat.

Bellingcat is a non-profit and the ability to carry out our work is dependent on the kind support of individual donors. If you would like to support our work, you can do so here. You can also subscribe to our Patreon channel here. Subscribe to our Newsletter and follow us on Twitter here and Mastodon here.

The post A Guide To Monitoring Conflict Amidst a Sea of Misinformation appeared first on bellingcat.

In June, Bellingcat ran 500 geolocation tests, comparing LLMs from various companies against each other, as well as Google Lens – a staple tool for finding the location of photos.

At the time, ChatGPT o4-mini-high emerged as the clear winner, with Google Lens outperforming most other models. Just two months later, with new versions of these AI tools available, we re-ran the trial – this time including Google “AI Mode,” GPT-5, GPT-5 Thinking, and Grok 4 into the mix.

The original test used 25 of Bellingcat’s own holiday photos. From cities to remote countryside, the images included scenes both with and without recognisable features – such as roads, signage, mountains, or architecture. Images were sourced from every continent.

For the updated trial, five test photos were excluded, as they had appeared in a previous article, thus compromising the integrity of the results.

All 24 models’ responses were ranked on a scale from 0 to 10, with 10 indicating an accurate and specific identification (such as a neighbourhood, trail, or landmark) and 0 indicating no attempt to identify the location at all.

Google AI Mode was shown to be the most capable geolocation tool overall. 

Grok 4 gave both better and worse answers compared to Grok 3 but, on average, scored marginally higher. However, it was still less accurate than older versions of Gemini and GPT. 

GPT-5, even in ‘Thinking’ and ‘Pro’ modes, was a considerable downgrade when compared with the capabilities demonstrated by GPT o4-mini-high. In one example, of a city street with skyscrapers in the background, o4-mini-high correctly identified the street, while GPT-5 in Thinking mode pointed to the wrong country. 

Support Bellingcat

Your donations directly contribute to our ability to publish groundbreaking investigations and uncover wrongdoing around the world.

Donate

Despite delivering faster answers, GPT-5 appeared to sacrifice accuracy. A surprising number of errors and a general sense of disappointment in the new model have also been reported by other users.

Bellingcat tested GPT-5 and its ‘Thinking’ mode via the Plus subscription, which costs roughly the same as access to 04-mini-high prior to its retirement. Five of the most difficult test images were also run through GPT-5 Pro. But even Pro, with a premium price tag of €200 per month, failed to geolocate the photos any more accurately than GPT 04-mini-high.

A Beach, a Hotel and a Ferris Wheel

The disparity between Google and the GPT models became even more apparent in Test 25 – a photo of a shoreline hotel in Noordwijk, the Netherlands, with a Ferris wheel rising just beyond the dunes.

In the previous trial, most older models – including those from GPT, Claude, Gemini and Grok – accurately identified the country as the Netherlands but failed to locate the town. Many latched onto the Ferris wheel but pointed instead to the seaside town of Scheveningen, which also has a Ferris wheel, though situated on a pier, not among the sand dunes.

However, the most recent models, GPT-5 Pro and Thinking, were even less accurate, identifying a beach in France – an entirely different country. 

Unfortunately for open source researchers, following the release of GPT-5, OpenAI removed the option to select older models such as o4-mini-high. After a wave of negative feedback, OpenAI reinstated GPT-4o as the default model for paid subscribers. However, the most capable geolocation models identified in Bellingcat’s testing remain inaccessible.

Google AI Mode, on the other hand, was the first, and only model so far, to correctly identify Noordwijk as the location in Test 25.  

Though AI Mode is powered by a version of Gemini 2.5, it outperformed Gemini 2.5 Pro Deep Research in these tests. Described by Google as its “most powerful AI search, with more advanced reasoning and multimodality,” AI Mode geolocated test images with greater accuracy than any GPT models, including our previous winner, o4-mini-high.

AI Mode is currently only available in India, United Kingdom and the United States.

The majority of models, at some point, returned a hallucination. Users should not rely solely on the answers provided by LLMs. Even the best options, including Google AI Mode, still, at times, confidently point to the wrong location. 

The difference in models’ capabilities compared with just two months ago shows how quickly this field is evolving. However, OpenAI’s recent changes also suggest that progress is not guaranteed, and that AI’s ability to geolocate may plateau or even worsen over time. As new models emerge, Bellingcat will continue to test them.

Thanks to Nathan Patin for contributing to the original benchmark tests.

Bellingcat is a non-profit and the ability to carry out our work is dependent on the kind support of individual donors. If you would like to support our work, you can do so here. You can also subscribe to our Patreon channel here. Subscribe to our Newsletter and follow us on Bluesky here and Instagram here.

The post LLMs Vs. Geolocation: GPT-5 Performs Worse Than Other AI Models appeared first on bellingcat.

Bellingcat’s Auto Archiver is a tool aimed at preserving online digital content before it can be modified, deleted or taken down. Publicly launched in 2022, it has preserved over 150,000 web pages and social media posts to date. The Auto Archiver has been used by Bellingcat’s journalists to preserve information on dozens of fast moving events such as the Jan. 6 riots – when we first used the tool internally – as well as gather digital evidence for our Justice and Accountability project and to monitor Civilian Harm in Ukraine.

The Auto Archiver has also been adopted by both large newsrooms and NGOs. It has been  used by individual researchers, journalists, activists, archivists, academics and developers as well.  With interest in the tool strong, we have worked hard to add to and improve it over time. But we have used the past few months to take a step back and to build a new and more robust ecosystem to further help individual organisations and researchers use and benefit from it.

Our aim has been to make it more reliable and even easier to use for more people. Today, we are happy to announce an updated version of the Auto Archiver which includes many new features like:

• Detailed documentation for all features and configurations
• A user-friendly interface designed for teams using a shared instance
• A new modular structure that improves the startup speed and reliability of the tool
• New features like chain of custody, perceptual hashing for deduplication, and techniques to avoid anti-bot measures and captchas on websites
• A user-friendly tool to configure the Auto Archiver, without the need to edit configuration text files

For an in-depth look at the changes made in this stable version of the Auto Archiver, see the What Changed, What Remains section further down in this article.

Automated Archiving and Collaboration – When to Use This Tool?

The latest version of the Auto Archiver has an easy-to-use web interface and a simplified installation process that makes it more straightforward to set up than before. However, some technical skills are still required for this initial process, and there are other tools available that could meet many of your archiving needs.

Support Bellingcat

Your donations directly contribute to our ability to publish groundbreaking investigations and uncover wrongdoing around the world.

Donate

If all you need is to archive a few unauthenticated URLs, we recommend using the Wayback Machine or Archive.today. Alternatively, WebRecorder’s browser extension ArchiveWebPage can create a replayable archive of a website you visit – even for content behind login walls. For batch processing, the Wayback Machine has a bulk upload service that accepts Google Sheets. If you individually need to record all your browser interactions and store content along the way there are paid options like Hunchly. Finally, if all you are interested in are videos and are comfortable with the command line, yt-dlp will probably be enough to download those, even in bulk.

But if you’re hoping to automate your archiving, or archive a large number of URLs in a collaborative environment, then this is where the Auto Archiver really shines. Its modular framework allows you or your team to customise archiving based on your needs, and provides a way to generate metadata that ensures others can trust that your archived content has not been tampered with. 

Learn more about what sites the Auto Archiver can archive here.

The Future of Web Archiving

Archiving the web is hard, especially when logins, captchas, and other bot prevention systems are in place. We will do our best to keep improving our Auto Archiver, but we note that it should be just one of many tools in your researcher’s toolkit. You can explore a variety of other useful tools in the Bellingcat Open Source Investigation Toolkit.

Still, if you want to support us on this journey of archiving critical information, you can:

• Download and use this tool
• Donate directly to Bellingcat
• Test, give feedback, and develop new features in our GitHub

For newsrooms:
If you work in a newsroom or research team and want to access a demo or help to deploy the Auto Archiver internally you can reach us at contact-tech@bellingcat.com with the Subject “Auto Archiver at [my team/organisation]” and tell us more about your organisation and archiving needs. Building a greater adoption base is the best way to ensure the future of this tool and its versatility.

What Changed, What Remains

Now that we have given a broad overview of the tool and its changes, what follows is a deeper look at how different parts of it work and interact. This will likely be of greater benefit for more technical users, and we again stress that successful users of the tool will likely need some technical knowledge to set it up for the first time. 

But help is available with our live Auto Archiver Documentation. This is where you will always find the latest information on how to install, configure or debug the tool. Even if some aspects mentioned in this article change in the coming years, the documentation will be your go-to space for the up to date instructions. 

If you have questions or problems please open an issue on GitHub. That’s where others will also be going to for help and constitutes our shared knowledge space.

A New Architecture

Many open source researchers, including at Bellingcat, favour using the Auto Archiver with the Google Sheets integration, which allows users to work collaboratively by adding links to a spreadsheet and letting the Auto Archiver run in the background. However, we have now made it simpler to integrate the Auto Archiver into other systems. One such example is ATLOS, a collaborative investigations platform that integrated the Auto Archiver and which has been used by Bellingcat and the Centre for Information Resilience. 

Integration is possible via the new modular architecture of the Auto Archiver and can be seen in the two new projects that we recently made public under open source code licenses: the Auto Archiver API and the Auto Archiver Web Interface.

Modules are the building blocks of the archiving pipeline and tell the tool how to run. They detail where to find the URLs, which archiving techniques to use, what additional processing to carry out on archived content and where and how to store it. Each module falls into a specific class:

1. Feeder modules specify where to read the URLs from. There’s one for Google Sheets, for example. 
2. Extractor modules download media and other metadata from a URL: our most versatile one is the Generic Extractor, which uses yt-dlp to download videos. However, extractors can be tailor made for specific platforms like the Telethon Extractor, which requires a Telegram account to download all media and metadata from the messages in public or private chats an account has joined. 
3. Enricher modules increase the value of the archived content with additional information or checks, such as hashing or timestamping the content for future consistency or chain of custody validations. 
4. Formatter modules collect and display the result of the process in a single formatted output. We use the HTML Formatter, as shown in this Bluesky post example.
5. Storage modules tell the tool where to put the files it downloaded or generated. The easiest is to store it locally. But to ensure better preservation the best practice is to use cloud storages like S3 or Google Drive. 
6. Database modules simply indicate where to save a record of this archive, such as whether archival was successful and which methods were used. You can use a CSV file and Google Sheets, for example. 

The modules documentation can be found here and it is there to help you understand how each module works and is configured. Configuring which modules to use is done via a YAML file. If you are not comfortable with those, we have you covered with a new interface called the configuration editor where you can visually create or edit your modules configuration. In fact, the first time you run the Auto Archiver a minimal working YAML configuration file is generated which you can use straight away to read URLs from the command line and store archived content locally.

Some platforms rate-limit or outright block IPs based on inauthentic behaviour. One of the strategies we employ to circumvent that is sending traffic through a proxy, which you can configure in specific modules like the Generic Extractor . We have been using Oxylab’s Residential Proxies as part of their Project 4beta successfully for over a year, but know that there are several good providers out there. 

If you are a developer, you can design new modules as needed using Python code, and we welcome it if you want to contribute those back to our code. Imagine a Feeder that is constantly scraping URLs from a Bluesky account, or an Enricher that uses an AI model to detect and blur graphic content. All of that is possible and easy to build in this new architecture. 

We hope you will enjoy the updated tool.

Please give us any feedback or suggestions for improvements by contacting us via contact-tech@bellingcat.com.

Bellingcat is a non-profit and the ability to carry out our work is dependent on the kind support of individual donors. If you would like to support our work, you can do so here. You can also subscribe to our Patreon channel here. Subscribe to our Newsletter and follow us on Bluesky here and Instagram here.

The post The Open Source Tool That Has Preserved 150,000 Pieces of Online Evidence appeared first on bellingcat.

Extreme weather events are increasingly forcing millions of people from their homes. Last year, floods and storms caused more people to be internally displaced around the world than war or violence. 

Driving food insecurity and competition for resources, climate change can also indirectly intensify conflict risks.

Analysing trends in both climate and conflict data is therefore not only critical for humanitarian response teams but also for journalists and researchers looking to understand the compounding impacts of weather-related hazards and violence on displaced communities.

In Part One of The Story of a Storm we showed you how to turn raw climate data into visualisations using NASA’s Giovanni tool and to plot it with Google Earth Pro Desktop.

In Part Two, we’ll focus on conflict and displacement data.

Using event data from the Armed Conflict Location & Event Data Project (ACLED) and geospatial information from OpenStreetMap (OSM), we will further analyse our case study from Part One – Cabo Delgado, a coastal province in the north of Mozambique. 

Cabo Delgado is a region besieged not only by cyclones but also by years of conflict. What began in 2017 as a series of localised attacks by Islamist militias has since escalated into an ongoing regional civil war between government forces and the Islamic State Central Africa Province (ISCAP). Close to a million people have been displaced, with extreme weather events including Cyclone Kenneth in 2019 and Cyclone Gombe in 2022 only adding to the unfolding humanitarian disaster. 

Step 1: Sourcing Conflict Incident Data 

Armed Conflict Location and Event Data (ACLED) provides real-time monitoring of global political violence and unrest by providing geo-referenced records of protests, riots, and violence targeting civilians. 

Most of ACLED’s data originates from public sources, including local media reports and social media posts, making it heavily dependent on the presence or absence of local journalism. This reliance can create gaps in the data. However, on occasion, ACLED may also gather intelligence from local partners on the ground. 

Each event entry includes key attributes such as dates, locations, involved actors, casualty figures and event descriptions. The data can be downloaded as CSV files with XY coordinates for spatial analysis. It is important to acknowledge that these coordinates often represent approximate locations, frequently defaulting to the nearest city centre rather than the precise location, which reduces spatial accuracy for detailed analysis. 

To download ACLED Data in CSV format, you will first need to create a free account. You will then need to generate an access key before navigating to the Data Export Tool. 

Given the complexity and longevity of the conflict in Cabo Delgado, with incidents dating back to 2017, we have chosen to split our data capture into three distinct time periods, described below. Each period will produce a separate CSV file, which, when uploaded to Google Earth Pro, will generate a separate data layer on the map.

Each time period reflects a shift in the level of violence and resulting displacement patterns of the community over time.

Period 1. Oct. 5, 2017 – March 23, 2021: Gradual Displacement
The first attacks begin in the northern districts. A steady flow of Internally Displaced Persons (IDPs) arrives further south, in and around Metuge. 

Support Bellingcat

Your donations directly contribute to our ability to publish groundbreaking investigations and uncover wrongdoing around the world.

Donate

Period 2. March 24, 2021 – Dec. 31, 2021: Peak Displacement
A large-scale attack on the town of Palma in March 2021 results in the displacement of more than 100,000 people.

Period 3. Jan. 1, 2022 – To Date: Post-Peak
Following the recapture of Palma and Mocimboa da Praia by Mozambican and allied forces, some IDPs begin to return to their homes.

To export the data navigate to the Data Export Tool and enter the information below, repeating this step for each time period:

Event Type: Battles, Explosions/Remote violence, Violence against civilians
Country: Mozambique
Date Range (From and To):
05/10/2017 – 23/03/2021 (Period 1: Gradual Displacement)
24/03/2021 – 31/12/2021 (Period 2: Peak Displacement)
01/01/2022 – To date (Period 3: Post-Peak)

Step 2: Mapping Conflict Data in Google Earth Pro

Launch Google Earth Pro on your desktop. In the top menu, click File → Import.

Navigate to one of the downloaded ACLED CSV files and select it. The pop-up window, Data Import Wizard, will appear. 

Select Field Type: Delimited. 

Choose Comma as the delimiter.

Click Next and leave the other settings as they are (ensure that the Latitude and Longitude fields are correctly selected). Click Finish to import the file. 

Step 3: Save and Style the Data

A new layer will appear below Temporary Places and Google Earth Pro will automatically display the imported points in the map. If they don’t appear, make sure the box next to the layer is ticked.

Right-click the dataset and choose Save Place As… to store it as a KML or KMZ file.

Repeat Steps 2 and 3 for all the ACLED CSV files you have downloaded. 

To change the icons or colours, right-click the layer file, Select Properties, then modify the Style/Color settings.

Step 4: Explore the Data

Click on any point to view details of the event, including the date, type, and location.

Be aware that ACLED may record multiple conflict incidents at the same location, so events can appear stacked on the map.

Toggle the time-period layers on and off to analyse spatio-temporal changes in conflict patterns. 

The visualisation reveals that during the first time period, Gradual Displacement (yellow), the violence was mostly concentrated in north-eastern Cabo Delgado, in and around Macomia, Mocimboa da Praia and Palma. In the second time period, Peak Displacement (red), incidents remained high in the north-east, leading to a large influx of IDPs into the Metuge region. By the third time period, Post-Peak (blue) the conflict had spread further south and west, increasing the risk of repeated displacements for populations already uprooted.

Zooming in on Displacement

To better understand localised displacement patterns, we can begin identifying specific IDP settlements of interest using high-resolution satellite imagery.

The Copernicus Emergency Mapping Service (EMS) conducted a satellite-based assessment of Cabo Delgado in 2021, identifying and mapping 27 IDP sites across the province. 

The EMS’ mission, to deliver near-real-time maps and geospatial data during or immediately after a critical event, aims to help responders on the ground make informed decisions quickly. Recent activations have included the wildfires in Portugal, floods in Germany, or post-landslide in Ethiopia. You can access their full database here. 

The EMS was activated over Cabo Delgado in response to the displacement of thousands of civilians after militants attacked and killed dozens in the coastal town of Palma. The map below shows seven of the IDP sites captured by the EMS in 2021 in the districts Metuge and Mecufi.

Based on the EMS data, we will focus on two out of the seven settlements: Cuaia (40°21’E 12°59’50″S)  and Ngalane (40°23’50″E 13°2’46″S). 

Copernicus EMS provides a detailed guide on what to look out for when analysing a potential IDP settlement, including housing types, tents, and other structures. Follow this link to download the PDF titled Technical Report.

Cuaia – A Temporary Settlement

As seen in the satellite imagery below, an organised camp including tents and sanitation facilities to the west of the small village of Cuaia was visible in May 2021. However, by May 2022, all the tents had disappeared, indicating that this was a temporary site which did not develop into a permanent IDP settlement.

Imagery of Cuaia (40°21’E 12°59’50″S) in May 2021 and May 2022. Source: GoogleEarthPro 2025

The tents’ light-coloured roofs and their uniform size and shape indicate that this settlement was temporary. Their neat arrangement in closely spaced rows suggests the camp was established by a humanitarian organisation. Just north of the tents, two rows of small dark structures are also visible, most likely sanitary facilities. Their alignment to the camp further supports the interpretation that this is a planned, non-permanent installation, distinct from the adjacent village to the east. In the southern part of the image, two buildings have appeared. A change in the colour and material of their roofs can be observed between 2021 and 2022, suggesting a shift from a temporary to more durable construction.

Ngalane – A Permanent Settlement

Located south of the city of Metuge, Ngalane’s expansion is clearly visible in satellite imagery – from a small community in 2020 to a larger, more established site by 2021. The visible increase in housing and agricultural fields suggests that Ngalane has transitioned from a temporary site into a permanent IDP settlement.

Imagery of Ngalane (40°23’50″E 13°2’46″S ) in August 2020 and May 2021

Rather than the organised, grid-like layout seen at temporary sites such as Cuaia above, the organic arrangement of dwellings in Ngalane, along with informal footpaths, small roads, and newly established small fields and cultivated plots, all indicate a level of permanence. While some buildings still have light-coloured roofs, indicating tent fabric, the overall spatial organisation of the settlement, combined with the presence of cultivated land, provides stronger evidence of a shift from short-term displacement to long-term settlement, reflecting efforts to rebuild livelihoods. 

Analysing the Settlements’ Digital Footprints 

Now that we’ve identified our two locations of interest, we can further analyse the settlements’ structures and their populations’ exposure to climate-related hazards, using building footprint data from OpenStreetMap (OSM). OSM is a collaborative project that provides free, crowdsourced geographic data, including information on roads, buildings, and other infrastructure. 

A settlement’s footprint data is based on the number, size, and shape of individual buildings, and can be used to estimate how many people live in a given area. Such data can be leveraged for exposure mapping, helping to identify a location, its density, and the resilience of its structures in relation to climate-related hazards such as floods, droughts, or cyclones.

Ngalane – an Incomplete Footprint

To see the current OSM building footprint of Ngalane, you’ll first need to create a free account. 

By navigating to the iD Editor window, we can see the OSM footprint of the original village (outlined in red), before the high influx of IDPs, has mostly been mapped. Red rectangles mark the buildings currently recorded in OSM. 

However, beyond the old village boundary, the newer settlements, home to the IDPs, remain entirely unmapped and therefore undocumented. In the event of a future storm or other climate-related hazard, the absence of mapping data for this community could result in their exclusion from impact assessments and early response planning.

Cuaia – an Out-of-Date Footprint

In Cuaia, 83 of the original village buildings have been mapped in OSM. None of the temporary structures have been documented, as they only briefly existed in 2021. However, when comparing the OSM data with more recent satellite imagery from 2025, it’s clear the settlement’s footprint is out of date: some buildings no longer exist, while others remain unmapped. 

The OSM metadata reveals that most structures were mapped between 2017 and 2019. If the data were needed for an accurate exposure assessment in the event of a cyclone or flood, for example, the OSM building footprints would need to be rapidly updated.

To address such time-sensitive scenarios, the initiative Humanitarian OpenStreetMap Team (HOT) organises rapid-response open mapping for crisis-affected areas. HOT is a global initiative that mobilises volunteers to help produce real-time geospatial information by tracing buildings, roads and rivers in OSM. Its Export Tool also allows users to select an area of interest and download custom extracts of OSM data, including building footprints, road networks, or land use, in various formats for geospatial analysis.

Analysing OSM data in Google Earth Pro

To overlay the settlement’s building footprints with the conflict and climate data, we will first export the relevant OSM layers using the HOT Export Tool, then import them into Google Earth Pro for analysis. 

You will need to create a free HOT account before taking the next steps. 

Step 1: Open the HOT Export Tool. Create a name for your dataset, then navigate to your area of interest, for example, Ngalane. Using the toolbar on the right, select the box tool and click and drag a rectangle over the area you want to export the data for.

Step 2: Under the Formats tab select Google Earth .kml then click Next.

Step 3: Tick the Buildings box to export all building footprint data for your selected area. Click Next.

Step 4: Click Create Export.

Step 5: Once the status shows Completed, click on the download link to save the data as a .zip file. 

Step 6: To import the data into Google EarthPro, either drag and drop the .kml file directly into the program or go to File -> Import -> then select the downloaded file. Once the data is imported, right-click on the layer in the panel on the left, select Properties, and under the Style, Color tab, customise the appearance of the building data on your map.

By visually comparing the distribution of mapped buildings to the satellite imagery in Google Earth Pro, you can assess the completeness and accuracy of OSM mapping in your area of interest. You can also explore historic satellite imagery through the Time-Tool to track changes in the building footprint over time and better understand how the settlement has developed.

Overlaying the Climate Data

In Part I of this guide, we mapped NASA climate data, specifically the accumulated rainfall brought by Cyclone Kenneth in 2019, to highlight the cyclone’s path and its impact on communities in Cabo Delgado. For a refresher on how to import climate data into Google Earth Pro, see Part I of this guide.

By overlaying the ACLED conflict incident data (2017 to 2021) with the accumulated rainfall data in Google Earth Pro, we can see how violence concentrated in the northwest was a major driver of displacement both before and during Cyclone Kenneth. However, displacement southward brought many directly into areas exposed to severe flooding.

Zooming in on our case study settlement, Ngalane, further reveals this community’s high exposure to the impacts of Cyclone Kenneth, including heavy rainfall and flooding.  

With at least 63 homes visible in Ngalane’s OSM building footprint, we can estimate a minimum number of residents at the time Cyclone Kenneth struck. However, since the footprint was incomplete, the actual population was likely higher. 

Mapping Cabo Delgado in 2025 

To date, numerous communities across Cabo Delgado remain unmapped and highly vulnerable to the next cyclone or flood, including Ngalane. 

Since the beginning of this year, more than 20,000 people have reportedly been displaced by ongoing violence, and these communities must be resettling somewhere. 

Recent satellite imagery of Ngalane reveals the rapid expansion of new buildings in all directions from the original village – all undocumented and therefore invisible on existing maps.

In order to update essential mapping data of the community, together with HOT we have created a task in the HOT OSM Tasking Manager to initiate mapping of new buildings in Ngalane. Follow this link to create an account, complete the OSM mapping tutorial and start mapping buildings in Ngalane! Your efforts will help generate essential mapping data for this settlement, supporting preparedness for future storms in the region.

To find out more about the intersection of climate, conflict, and displacement, explore the Story Map Climate, disasters and conflict in Cabo Delgado by the Red Cross Red Crescent Climate Centre (2022). It offers a retrospective analysis of the impacts of Cyclones Kenneth and Idai, framed within the context of conflict and displacement.

Bellingcat is a non-profit and the ability to carry out our work is dependent on the kind support of individual donors. If you would like to support our work, you can do so here. You can also subscribe to our Patreon channel here. Subscribe to our Newsletter and follow us on Bluesky here and Mastodon here.

The post The Story of a Storm Part II: Visualising Conflict and Displacement Data appeared first on bellingcat.

An ambiguous city street, a freshly mown field, and a parked armoured vehicle were among the example photos we chose to challenge Large Language Models (LLMs) from OpenAI, Google, Anthropic, Mistral and xAI to geolocate. 

Back in July 2023, Bellingcat analysed the geolocation performance of OpenAI and Google’s models. Both chatbots struggled to identify images and were highly prone to hallucinations. However, since then, such models have rapidly evolved. 

To assess how LLMs from OpenAI, Google, Anthropic, Mistral and xAI compare today, we ran 500 geolocation tests, with 20 models each analysing the same set of 25 images. 

Our analysis included older and “deep research” versions of the models, to track how their geolocation capabilities have developed over time. We also included Google Lens to compare whether LLMs offer a genuine improvement over traditional reverse image search. While reverse image search tools work differently from LLMs, they remain one of the most effective ways to narrow down an image’s location when starting from scratch.

The Test

We used 25 of our own travel photos, to test a range of outdoor scenes, both rural and urban areas, with and without identifiable landmarks such as buildings, mountains, signs or roads. These images were sourced from every continent, including Antarctica. 

The vast majority have not been reproduced here, as we intend to continue using them to evaluate newer models as they are released. Publishing them here would compromise the integrity of future tests.

Each LLM was given a photo that had not been published online and contained no metadata. All models then received the same prompt: “Where was this photo taken?”, alongside the image. If an LLM asked for more information, the response was identical: “There is no supporting information. Use this photo alone.”

We tested the following models:

This was not a comprehensive review of all available models, partly due to the speed at which new models and versions are currently being released. For example, we did not assess DeepSeek, as it currently only extracts text from images. Note that in ChatGPT, regardless of what model you select, the “deep research” function is currently powered by a version of o4-mini. 

Support Bellingcat

Your donations directly contribute to our ability to publish groundbreaking investigations and uncover wrongdoing around the world.

Donate

Gemini models have been released in “preview” and “experimental” formats, as well as dated versions like “03-25” and “05-06”. To keep the comparisons manageable, we grouped these variants under their respective base models, e.g. “Gemini 2.5 Pro”. 

We also compared every test with the first 10 results from Google Lens’s “visual match” feature, to assess the difficulty of the tests and the usefulness of LLMs in solving them. 

We ranked all responses on a scale from 0 to 10, with 10 indicating an accurate and specific identification, such as a neighbourhood, trail, or landmark, and 0 indicating no attempt to identify the location at all.

And the Winner is…

ChatGPT beat Google Lens.

In our tests, ChatGPT o3, o4-mini, and o4-mini-high were the only models to outperform Google Lens in identifying the correct location, though not by a large margin. All other models were less effective when it came to geolocating our test photos.

We scored 20 models against 25 photos, rating each from 0 (red) to 10 (dark green) for accuracy in geolocating the images.

Even Google’s own LLM, Gemini, fared worse than Google Lens. Surprisingly, it also scored lower than xAI’s Grok, despite Grok’s well-documented tendency to hallucinate. Gemini’s Deep Research mode scored roughly the same as the three Grok models we tested, with DeeperSearch proving the most effective of xAI’s LLMs.

The highest-scoring models from Anthropic and Mistral lagged well behind their current competitors from OpenAI, Google, and xAI. In several cases, even Claude’s most advanced models identified only the continent, while others were able to narrow their responses down to specific parts of a city. The latest Claude model, Opus 4, performed at a similar level to Gemini 2.5 Pro. 

Here are some of the highlights from five of our tests.

A Road in the Japanese Mountains

The photo below was taken on the road between Takayama and Shirakawa in Japan. As well as the road and mountains, signs and buildings are also visible.

Gemini 2.5 Pro’s response was not useful. It mentioned Japan, but also Europe, North and South America and Asia. It replied:

“Without any clear, identifiable landmarks, distinctive signage in a recognisable language, or unique architectural styles, it’s very difficult to determine the exact country or specific location.”

In contrast, o3 identified both the architectural style and signage, responding:

“Best guess: a snowy mountain stretch of central-Honshu, Japan—somewhere in the Nagano/Toyama area. (Japanese-style houses, kanji on the billboard, and typical expressway barriers give it away.)”

A Field on the Swiss Plateau

This photo was taken near Zurich. It showed no easily recognisable features apart from the mountains in the distance. A reverse image search using Google Lens didn’t immediately lead to Zurich. Without any context, identifying the location of this photo manually could take some time. So how did the LLMs fare?

Gemini 2.5 Pro stated that the photo showed scenery common to many parts of the world and that it couldn’t narrow it down without additional context. 

By contrast, ChatGPT excelled at this test. o4-mini identified the “Jura foothills in northern Switzerland”, while o4-mini-high placed the scene ”between Zürich and the Jura mountains”.

These answers stood in stark contrast to those from Grok Deep Research, which, despite the visible mountains, confidently stated the photo was taken in the Netherlands. This conclusion appeared to be based on the Dutch name of the account used, “Foeke Postma”, with the model assuming the photo must have been taken there and calling it a “reasonable and well-supported inference”.

An Inner-City Alley Full of Visual Clues in Singapore

This photo of a narrow alleyway on Circular Road in Singapore provoked a wide range of responses from the LLMs and Google Lens, with scores ranging from 3 (nearby country) to 10 (correct location).

The test served as a good example of how LLMs can outperform Google Lens by focusing on small details in a photo to identify the exact location. Those that answered correctly referenced the writing on the mailbox on the left in the foreground, which revealed the precise address.

While Google Lens returned results from all over Singapore and Malaysia, part of ChatGPT o4-mini’s response read: “This appears to be a classic Singapore shophouse arcade – in fact, if you look at the mailboxes on the left you can just make out the label ‘[correct address].’”  

Some of the other models noticed the mailbox but could not read the address visible in the image, falsely inferring that it pointed to other locations. Gemini 2.5 Flash responded, “The design of the mailboxes on the left, particularly the ‘G’ for Geylang, points strongly towards Singapore.” Another Gemini model, 2.5 Pro, spotted the mailbox but focused instead on what it interpreted as Thai script on a storefront, confidently answering: “The visual evidence strongly suggests the photo was taken in an alleyway in Thailand, likely in Bangkok.”  

The Costa Rican Coast

One of the harder tests we gave the models to geolocate was a photo taken from Playa Longosta on the Pacific Coast of Costa Rica near Tamarindo. 

Gemini and Claude performed the worst on this task, with most models either declining to guess or giving incorrect answers. Claude 3.7 Sonnet correctly identified Costa Rica but hedged with other locations, such as Southeast Asia. Grok was the only model to guess the exact location correctly, while several ChatGPT models (Deep Research, o3 and the o4-minis) guessed within 160km of the beach.

An Armoured Vehicle on the Streets of Beirut

This photo was taken on the streets of Beirut and features several details useful for geolocation, including an emblem on the side of the armored personnel carrier and a partially visible Lebanese flag in the background. 

Surprisingly, most models struggled with this test: Claude 4 Opus, billed as a “powerful, large model for complex challenges”, guessed “somewhere in Europe” owing to the “European-style street furniture and building design”, while Gemini and Grok could only narrow the location down to Lebanon. Half of the ChatGPT models responded with Beirut. Only two models, both ChatGPT, referenced the flag.  

So Have LLMs Finally Mastered Geolocation?

LLMs can certainly help researchers to spot the details that Google Lens or they themselves might miss.

One clear advantage of LLMs is their ability to search in multiple languages. They also
appear to make good use of small clues, such as vegetation, architectural styles or signage. In one test, a photo of a man wearing a life vest in front of a mountain range was correctly located because the model identified part of a company name on his vest and linked it to a nearby boat tour operator.

For touristic areas and scenic landscapes, Google Lens still outperformed most models. When shown a photo of Schluchsee lake in the Black Forest, Germany, Google Lens returned it as the top result, while ChatGPT was the only LLM to correctly identify the lake’s name. In contrast, in urban settings, LLMs excelled at cross-referencing subtle details, whereas Google Lens tended to fixate on larger, similar-looking structures, such as buildings or ferris wheels, which appear in many other locations.

Heat map to show how each model performed on all 25 tests

Enhanced Reasoning Modes

You’d assume turning on “deep research” or “extended thinking” functions would have resulted in higher scores. However, on average, Claude and ChatGPT performed worse. Only one Grok model, DeeperSearch, and one Gemini, Gemini Deep Research, showed improvement. For example, ChatGPT Deep Research was shown a photo of a coastline and took nearly 13 minutes to produce an answer that was about 50km north of the correct location. Meanwhile, o4-mini-high responded in just 39 seconds and gave an answer 15km closer.  

Overall, Gemini was more cautious than ChatGPT, but Claude was the most cautious of all. Claude’s “extended thinking” mode made Sonnet even more conservative than the standard version. In some cases, the regular model would hazard a guess, albeit hedged in probabilistic terms, whereas with “extended thinking” enabled for the same test, it either declined to guess or offered only vague, region-level responses.

LLMs Continue to Hallucinate

All the models, at some point, returned answers that were entirely wrong. ChatGPT was typically more confident than Gemini, often leading to better answers, but also more hallucinations. 

The risk of hallucinations increased when the scenery was temporary or had changed over time. In one test, for instance, a beach photo showed a large hotel and a temporary ferris wheel (installed in 2024 and dismantled during winter). Many of the models consistently pointed to a different, more frequently photographed beach with a similar ride, despite clear differences.

Final Tips

Your account and prompt history may bias results. In one case, when analysing a photo taken in the Coral Pink Sand Dunes State Park, Utah, ChatGPT o4-mini referenced previous conversations with the account holder: “The user mentioned Durango and Colorado earlier, so I suspect they might have posted a photo from a previous trip.” 

Similarly, Grok appeared to draw on a user’s Twitter profile, and past tweets, even without explicit prompts to do so. 

Video comprehension also remains limited. Most LLMs cannot search for or watch video content, cutting off a rich source of location data. They also struggle with coordinates, often returning rough or simply incorrect responses. 

Ultimately, LLMs are no silver bullet. They still hallucinate, and when a photo lacks detail, geolocating it will still be difficult. That said, unlike our controlled tests, real-world investigations typically involve additional context. While Google Lens accepts only keywords, LLMs can be supplied with far richer information, making them more adaptable.

There is little doubt, at the rate they are evolving, LLMs will continue to play an increasingly significant role in open source research. And as newer models emerge, we will continue to test them. 

Infographics by Logan Williams and Merel Zoet

The post Have LLMs Finally Mastered Geolocation? appeared first on bellingcat.

This guide is part of a collaboration between Bellingcat and Evident on detecting AI-generated products. You can watch Evident’s video here. 

Sipping coffee from a mug carved from mineral rock, its surface glimmering with amethyst, rose quartz and other crystals, sounds almost too magical to be real.

And unfortunately, as some shoppers discovered, it was. 

Ads for crystal coffee cups, like the one shown in the Facebook post below, have appeared across the internet. The artisan mugs, available in color variations like blue, green and pink, were being sold on a swathe of platforms ranging from independent Facebook pages to large retailers including Amazon. However, when customers received these mugs, they were in for a surprise.

In the comment section of this Facebook post, users shared images of the mugs they had bought, which bore little resemblance to the fantastical images in the listing. 

Major advancements in artificial intelligence in recent years have made it harder to differentiate between what’s real or fake, not just when it comes to photos and videos of people, but also in product listings. 

There has also been an increase in AI-generated books being sold on platforms including Amazon, with some even showing up in libraries without any disclosure that they are AI-generated.  

But AI is not perfect, and if you look closely, you can often detect several tell-tale signs of a fake. In this guide, we walk you through some questions that savvy shoppers can ask to identify “red flags”, using just critical thinking and basic investigative tools such as reverse image searches. 

Does the Image Make Sense?

Many AI-generated images have some sort of “sheen” or look to them that can set off alarm bells.

Take this image of one of the “crystal coffee mugs”. At first glance, it looks like a beautiful mug. But if you look closer, you might notice defects in the image. 

There are multiple areas on this mug where the lines of the “crystal” do not align. These broken or inconsistent lines are red flags. There also appears to be some sort of defect in the centre of the mug that resembles a smudge from digital painting, rather than a natural flaw in rock or crystal. The blurriness of the smudge is a feature often seen in AI-generated images. And at the top of the rim, a section of the mug begins to fade out and disappear, suggesting the image has been manipulated.  

It is also useful to think about how this product would work in practice. For example, the mug shown in the picture below appears to be made of some sort of lava-type stone, with glowing red light emanating from the cracks. The lighting on the mug appears to be artificial and since there are no visible wires, it would probably need a battery to power it. However, the listing does not specify whether a power source is required or included, which should raise suspicions that the image could be AI-generated. 

Are There Multiple Angles and Pictures of the Product?

AI image generators can create convincing images, but they are not great at producing the same image consistently. Authentic listings will often show the same item from multiple angles so customers can see what the object looks like before purchasing. If you only see one photo of the item, that is a red flag that the listing may be using an AI-generated image. Sellers will often take one amazing photo and place the object in multiple “scenes” but you may notice that they don’t show any other angle of the item.

Where there are multiple photos of the product, it is also worth considering whether it looks like the same product in all of the pictures. In the below Etsy listing for a crocheted Highland cow pattern, for example, there are multiple photos of crocheted Highland cows, but they are not consistently the same pattern or design. 

Given that this listing is supposed to be for a pattern – a template which crocheters can follow to create the product shown – it’s suspicious that there are different colours, shapes and materials used in these photos. It may mean that these images were created by an image generator that was not able to replicate the exact same stuffed cow. 

You can also focus on differences in small details, like the placement of the nostril holes, between the photos or even sometimes in the same photo. For example, in the very first image of this listing, the two nostril holes are slightly different shapes. In the subsequent images, there are slight variations in how these nostril holes are depicted on the cows. The shape of the horn, body, hooves, and the scarf, all have variations between the images.

The image also does not make sense as, if you’re familiar with crochet, you may notice that the hairs shown on the head, body and legs of some of these sample stuffed cows do not match the texture of the type of yarn typically used for crochet. In one of the seller’s replies to customer complaints, they even confirmed that these images were AI-generated. 

Bonus Tip: Zooming In On Eyes

One useful hint for identifying an AI-generated image of a human face is to look at the reflective light in the eyes of the people shown and see if there are any abnormal patterns. The same principles may also be applied to images of animals or products such as this Highland cow, where there are shiny surfaces that reflect light. 

In real photos, the shape of the light being reflected in the eyes is typically identical or nearly identical if they are facing the same direction. But AI-image generators have not perfected this feature. In some of the pictures of the stuffed cow’s eyes, the shape of the light differs. Since the stuffed toy depicted is facing forward, the light in both eyes should be the same as they should be reflecting back the same light source. 

However, as you can see from the photos below, they are different – another indicator that these images are AI-generated. 

Have You Thoroughly Read the Listing?

To protect themselves from being accused of misrepresentation or having their listings taken down by platforms, some sellers may hide disclosures within text or images – relying on the assumption that buyers may not closely read the details of their listings. If called out, they may then claim that the images were only for illustration.

To avoid falling prey to such tactics, look out for specific product details including the materials and dimensions, and whether the seller discloses that they are using AI-generated photos. Sellers may include these details within the descriptions to indicate that the item will be different to what is depicted in the listing’s image.

In the case of the mugs, some listings state that the product is a “crystal-like” or “mineral-inspired” design, indicating that it is made from another type of material.   

In another example, these cute animal-themed ornaments look three-dimensional and hyper-realistic in their eBay listing. However, if you look at the full listing title, the seller specifies that the ornament is two-dimensional – in other words, flat. 

In one post on Reddit, a user said they purchased animal ornaments that looked similar to those in this listing. Based on their post, the ornaments they received were flat acrylic discs with images printed on them. If – as in the listing we saw – the seller indicated this in some way in the title or description, buyers who purchased the ornaments based on the AI-generated images alone may find it harder to seek recourse. 

Scammers rely on people making impulse purchases. Being consistent in reading the details can protect you from these surprises. 

Are There Pictures Posted by Buyers?

Fake reviews can occur on just about any platform, not just retail websites. There are whole networks dedicated to creating fake reviews, so we cannot just rely on positive customer feedback to determine if a product is trustworthy to purchase.

However, it takes more effort to create a fake review that includes images of the product.

In addition to reading the text reviews, look to see if anyone has posted photos of the item in question, and compare how it looks to the item shown in the listing. Are these photos just taken from the listing? Do they look like they’re showing the same item? Does the background or setting look like someone’s home?

If the item is on a platform that does not have reviews, or seems to be a newly listed item without many reviews, you can try doing a reverse image search of the item. Some sellers will take down a listing when they start getting negative reviews, and then relist the item again. Reverse image searches may pick up archived versions of these older listings, where you might be able to find negative user reviews. You may even find multiple sellers listing the same item and see negative reviews for the product from another seller. 

If the same product is being sold elsewhere, check the photos and customer reviews. If you cannot find any photos of the item other than those supplied by the seller, you may want to investigate further. 

Is It Too Good to Be True?

Many of us often search for the cheapest items online, looking to get the most bang for our buck, but it is also important to be alert to deals that seem too good to be real. 

It is always good practice to compare prices. Is there a difference of a few dollars, or a few hundred? Should a highly intricate sweater be this cheap?

Reverse image searches can also be a good tool to use here. If there is a drastic difference in cost on listings showing similar products, this should give you pause. While there may be legitimate reasons for such price differences, such as the country of origin or bulk buying by sellers, this could also indicate the lower priced listings are a scam. 

For example, the below stained glass lamp in the shape of a cat is sold on Walmart, Amazon, and eBay, among others, for under US$23.

An online search for other stained glass lamps returns listings of lamps in more simple domed shapes that cost at least a hundred US dollars – significantly more than the cat-shaped ones that could be more costly to produce because of their intricacy. 

Additionally, a Google search for cat-related stained glass lamps returned images of lamps in simpler, boxy shapes. Other than the images that matched the cat lamp above, the rest of the results were all listed at a significantly higher price point than US$25. From this search, it appears that not only is the price point too low but that even the design and shape of the cat lamp is not representative of other lamps available on the market.

While it’s possible that an item that looks somewhat similar may have a large price difference due to branding or geographical origin, this discrepancy is a potential red flag that you might want to investigate further to see if this is a real deal or a scam.

In this case, a reverse image search of this lamp showed that some customers who ordered it received a cheap plastic item with airbrushed paint rather than stained glass, and nothing at all close to the object in the listing. 

Who Is Behind This Item? 

Finally, it’s always helpful to look beyond the listing and consider who is profiting from the purchase. 

Is this a name-brand item, or does it appear to be some mysterious seller that has popped up overnight? Does this seller have a website, or do they only exist on Facebook or Amazon? Is the seller’s account brand new? If they sell other items, what do the customer reviews say? Is the seller using an AI-generated image as their profile picture?

If it is a book you are purchasing, look to see who the author is. Does this person exist? Is there a legitimate publisher of the book, or is it just through Amazon’s self-publishing? While self-publishing does not automatically mean a book is untrustworthy, this process has fewer checks and balances compared with books from publishers, which usually go through a review before being released. 

Doing an online search on the author should reveal other information about them. If they are real, does it appear that this book was authored by them? Do they promote it on other platforms? Could someone else be using their name without permission? 

For example, the cookbook shown below lists the author as “Ethan Neulife”. Their bio on Amazon describes an “experienced author” but a search of his name did not turn up anything except cached listings on Amazon’s various online marketplaces for this book, which are no longer accessible. There were no social media profiles or personal sites under this name, articles about his books or any contact information available online. 

A reverse image search of the profile picture using several different search engines did not return any exact matches except those on Amazon. However, it did suggest stock images that were marked as AI-generated. 

For example, one of the suggested results came from a website that specialises in AI-generated stock imagery, FreePik. The image titled “Portrait of Businessman on White Background”, generated using Midjourney 6, bore a striking resemblance to the profile photo of “Ethan Neulife”.

A reverse image search of a profile picture that returns results from AI-image websites may indicate that the image you are searching for is also AI-generated.

These kinds of checks can be done before purchase to ensure that the book or product you are buying comes from a legitimate company or person. There are, of course, always small businesses or individuals vigilant about online privacy who may not have much information about themselves online, but if in doubt, it is a good idea to do some basic online research on the seller to get a sense if they are legitimate or a potential scammer.

Bellingcat is a non-profit and the ability to carry out our work is dependent on the kind support of individual donors. If you would like to support our work, you can do so here. You can also subscribe to our Patreon channel here. Subscribe to our Newsletter and follow us on Twitter here and Mastodon here.

The post Don’t Get Scammed! Tips For Spotting AI-Generated Fake Products Online appeared first on bellingcat.