Sign up here to receive Bellingcat’s biggest investigations by email as soon as they are published. Become a member and support our work here.

An online ad for Ben McAdams, the Democratic candidate for Congress in Utah, emphasised that he could be trusted because he “fought back against MAGA”.

In Texas, another ad promoted Republican candidate Chris Gober, calling him a “MAGA warrior”.

What’s interesting is that both ads were ultimately paid for with money from the same organisation – Leading the Future (LTF), a Super PAC which opposes regulations that “constrain” innovation in AI. The network has received US$125 million in funding in total from OpenAI president Greg Brockman, his wife Anna Brockman, tech-focused venture capital firm Andreessen Horowitz (a16z) and a16z’s founders, Federal Election Commission (FEC) filings show. 

Directly opposing LTF is Public First Action (PFA), which calls for AI safety regulations and guardrails. PFA has received $40 million in funding from AI company Anthropic. 

Like the ads run by organisations funded by LTF, the Super PACs funded by PFA also ran ads that appeared supportive or critical of Trump depending on the voters they were targeting. 

None of these ads mentions AI.

As the US 2026 midterm elections approach on Nov. 3, Super PAC spending on political advertising has reportedly soared to record highs, surpassing US$11 billion last week and exceeding the amount spent on any other election cycle to date. 

Super PACs are a particular type of political action committee (PAC) that may receive unlimited contributions from individuals, corporations, labour unions and other PACs for the purpose of financing independent expenditures and other independent political activity.

In this year’s elections, LTF’s and PFA’s Super PACs, funded by the AI industry heavyweights, have played an outsized role in supporting or opposing candidates through advertising. But as polls show distrust of AI crosses party lines, the advertisements that Silicon Valley’s leaders are paying millions of dollars to try to sway the results in their chosen candidates’ favour often do not mention AI at all. 

While LTF is a registered Super PAC itself, FEC filings show it mainly disburses money to other Super PACs in its network which run political ads. 

Of about $52 million that LTF has disbursed to date, $20 million went to Think Big, which targets the Democratic voter base, and another $20 million went to American Mission, which focuses on Republicans. 

PFA, meanwhile, is registered as a “social welfare”, or 501(c)(4), organisation. This type of non-profit is not required to disclose its donors or disbursements. However, PFA contributes funds to three Super PACs: the bipartisan Public First PAC, Jobs and Democracy for Democratic candidates and Defending Our Values for Republicans. FEC filings for these Super PACs show Public First PAC received $1 million, Jobs and Democracy received $15.4 million and Defending Our Values received $7.9 million from PFA. Public First PAC’s FEC filings also show that $4 million of its $4.4 million in disbursements disclosed so far went to Jobs and Democracy, while another $300,000 went to Defending Our Values. 

Super PACs are allowed to accept unlimited contributions from other political committees, and other similar networks of Super PACs have been observed in the current election cycle, such as cryptocurrency-linked Fairshake and its affiliated Super PACs.

Bellingcat analysed data on more than 300 online video ad campaigns paid for by Think Big, American Mission, Jobs and Democracy and Defending Our Values from Google’s Ad Transparency Center. These ad campaigns, many of which used the same videos in slightly different formats or across different platforms, supported or opposed 40 candidates. 

Only ads supporting or opposing two candidates mentioned AI, and none mentioned data centres, a top issue this election cycle. A New York Times analysis of 95 video ads paid for by the same four Super PACs, using data gathered by media tracking firm AdImpact, had similar findings. FEC rules state that Super PACs can spend money to support or oppose a candidate, but cannot coordinate with the candidate or their campaign on this. There is no evidence that any of the candidates or campaigns for which the Super PACs ran ads have coordinated with them.

Michael Beckel, Director of Money in Politics Reform at cross-partisan political reform group Issue One, told Bellingcat that voters are at a disadvantage when special interests fund Super PACs running ads that don’t mention the priority issues of those underwriting them. 

“Special interests seem keen on moving voters with messaging around hot-button issues, all while advancing a specific policy agenda that viewers of ads are kept in the dark about,” he said. The result of this, Beckel added, was that “the industry gains a stronger hand in Washington, while voters get little chance to judge the agenda their votes could help advance”. 

A representative for PFA stated: “We endorse officials who support regulating AI. The aligned PACs run ads on the issues that matter most immediately to voters in each race, and for most electorates right now that means the economy, affordability, public safety, and other issues that they are facing on a daily basis.” 

Jobs and Democracy, Defending Our Values, LTF, Think Big and American Mission did not respond to questions about their political advertising as of publication. 

A recent survey by think tank Data for Action and non-profit Guardrails Action found that about three-quarters of likely voters would be less inclined to vote for a candidate if they found out they had accepted a donation from an AI Super PAC. 

The ads distributed by the AI-backed Super PACs supporting or opposing 29 candidates mentioned economic issues such as taxes, the cost of living, consumer rights and local jobs or small businesses. Another 16 mentioned immigration or law enforcement and 17 mentioned US President Donald Trump or his “Make America Great Again” (MAGA) movement. 

Like the ads mentioning Trump, those on immigration were sharply divided in messaging depending on whether they were targeting Democratic or Republican voters, even when they were paid for by the AI Super PACs in the same network. 

LTF’s Think Big sponsored an ad for Democratic candidate Jeremy Moss in Michigan that talks about how he stood up to Immigration and Customs Enforcement (ICE), while American Mission’s ad for Republican Clay Fuller in Georgia mentions how he is focused on securing the southern border. 

We found similar messaging in ads mentioning immigration that were paid for by Super PACs in PFA’s network. Defending Our Values ran an ad in support of Republican John Cowan in Georgia emphasising that he will combat illegal immigration, whereas Jobs and Democracy’s ad for Democrat Valerie Foushee in North Carolina said she was “leading the fight to hold ICE accountable”.

For the ads that did mention AI, almost all were in support of Alex Bores’ primary in New York. PFA’s Jobs and Democracy PAC ran 152 ads supporting Bores’ candidacy, at least some of which mentioned that he was in favour of safeguards on AI and holding Big Tech accountable. 

LTF’s Think Big, on the other hand, ran 22 ads opposing Bores’ candidacy. While they did not mention his stance on AI, many of Think Big’s ads focused on him receiving funding from Anthropic, “dark money Super PACs” and his former employer Palantir, the software giant whose technology, including a new AI system, is used in ICE’s immigration enforcement operations. 

Jobs and Democracy’s advertiser page also showed two ads for Josh Gottheimer in New Jersey that focused on the threats of AI scams and AI-generated media, urging voters to call Gottheimer to tell him to “stand strong for AI safeguards”. 

Major donors to LTF that Bellingcat reached out to – Greg Brockman, a16z, and a16z founders Marc Andreessen and Ben Horowitz – did not respond to requests for comment on their funding to the Super PACs as of publication. We could not find any publicly available contact information for Anna Brockman, but we also asked about her contributions to LTF in questions to her husband.

OpenAI did not respond to questions about the Brockmans’ contributions to LTF as of publication. However, the company published a statement on the topic in June, which said “OpenAI does not direct the activities of LTF, or have visibility into their operations” and that “any engagement with that organisation has been in a personal capacity, not on behalf of the company”. 

PFA’s main public donor, Anthropic, did not respond to Bellingcat’s queries about its funding for PFA and PFA’s political advertising as of publication. It said in a statement on its website in July that its donations to PFA were “to help promote policies that will maintain meaningful safeguards, sustain America’s AI leadership, and demand transparency from the companies building the most powerful AI models”.

Unknown Vendors

Federal campaign finance laws require federal political committees to file periodic campaign finance reports disclosing their receipts and disbursements. 

Our analysis of FEC filings for LTF and the Super PACs in its network revealed that they were channelling more than $36.4 million in total towards four companies that were newly set up, with no apparent online presence outside of these filings. 

A complaint to the FEC by the non-profit Campaign Legal Center (CLC) in May cited American Mission and Think Big’s payments to Summit Ridge and Lantern Productions, alleging that the Super PACs had “unlawfully evaded federal reporting requirements by funnelling their payments for goods and services through shell companies set up to conceal how the super PACs are actually spending their money”. The FEC has not yet ruled on the complaint. LTF, American and Think Big did not respond to Bellingcat’s queries about the complaint as of publication.

FEC filings show that LTF paid a company called Prusik Partners LLC more than $966,000 for operational expenditures for a range of services including communications and strategy consulting, as well as legal fees. 

The network’s Democrat-focused Super PAC, Think Big, paid a company called Lantern Production Consultants LLC more than $20 million, based on filings for disbursements from Aug. 20, 2025 to Oct. 14, 2026. 

The payments included $3.17 million for operational expenditures for a range of services such as media consulting and research, and another $17 million for independent expenditures. The FEC defines an independent expenditure as one that pays for “a communication that expressly advocates for the election or defeat of a clearly identified candidate”. The highest independent expenditures paid to Lantern ($8 million) were to oppose Bores’ candidacy, the filings show.

Think Big also declared in its filings that it had paid a company called Blue Horizon Compliance Consultants LLC almost $55,700 for “compliance consulting”. 

Republican-focused American Mission, on the other hand, paid a company called Summit Ridge Media LLC more than $13 million, based on filings for disbursements from Aug. 20, 2025 to Oct. 14, 2026. 

This included $2.2 million for operating expenditures such as polling and website services, and $10.9 million in independent expenditures, with the highest amount (about $960,400) going towards communications in support of James Morris Kingston, the Republican candidate for the House of Representatives in Georgia.

Bellingcat found that all four of these LLCs listed mailing offices and office supply goods stores in Nevada as their mailing addresses. We could not find records of them doing business with any entity outside of the LTF network – in fact, they appeared to have no online presence aside from their names appearing in the filings and business registrations. 

Despite their addresses being listed in Nevada, none of the companies could be found in the state’s public business registry. Instead, Bellingcat found registrations for Prusik, Lantern Production and Summit Ridge across the country – in Delaware. 

While Delaware is the second-smallest state in the US, more than half of all Fortune 500 companies are incorporated there according to the Delaware Secretary of State. The state’s business-friendly laws mean companies can reportedly incorporate in minutes and are only required to name a representative – who can be a paid agent unrelated to the business – in public registration documents, not the beneficial owners.

Articles of incorporation from the Delaware Secretary of State showed that Prusik was registered on July 22, 2025, while Lantern Production and Summit Ridge were both registered on Oct. 30, 2025. We also found a Delaware registration for a company called “Blue Horizon Compliance Group LLC”, registered on Oct. 30, 2025. While the name differs slightly from the Blue Horizon Compliance Consultants LLC that Think Big declared payments for, the registration was signed by the same agent listed in the documents for Lantern Production and Summit Ridge. 

The filings show that all of these companies – including Blue Horizon Compliance Group, if this was indeed the same entity referred to in Think Big’s filings – began receiving payments from the Super PACs within two to three months of their business registrations.

Saurav Ghosh, CLC’s Director of Federal Campaign Finance Reform, told Bellingcat that the group filed the complaint because “when groups funded by big corporate special interests evade transparency laws, voters lose”. 

“Federal reporting requirements help ensure transparency in our elections and protect voters’ legal right to know where the money is going,” Ghosh said. “Secret spending schemes like this make it virtually impossible to ensure that our elected officials are staying accountable to voters instead of wealthy special interests.”

We were unable to reach out to the LLCs as none of them had phone numbers or email addresses listed online. LTF, Think Big, and American Mission did not respond to questions about the LLCs, including whether “Blue Horizon Compliance Consultants LLC” listed in Think Big’s filings was the same company as Delaware-registered “Blue Horizon Compliance Group LLC”. However, a WIRED story on these same LLCs quoted a spokesperson from LTF as saying that the complaint was “a desperate ploy for attention from Public First, Anthropic, and the groups aligned with their outside organisations”. WIRED also quoted a Public First spokesperson denying that the group had anything to do with the complaint.  

Beckel suggested that stricter transparency measures relating to Super PAC spending would help inform and empower voters. For instance, Congress could pass legislation to require ads to include the names of the top donors, he said, adding that similar rules had been passed with cross-partisan support in places like Alaska and Arizona. 

“Congress would be doing voters a huge service by passing more robust transparency requirements for political ads. Voters deserve to know who is behind the ads they’re seeing,” Beckel said. 

About the Data

Google Ads Analysis

We reviewed ads available through Google’s Ads Transparency Center. These are ads served on Google platforms like Search, Display, Gmail and YouTube. We filtered specifically for ads shown between Jan. 1 and Oct. 23, 2026 in the US. 

We focused only on video ads classified as political advertisements, not text- or image-based ads. We looked at ads purchased by the following advertisers: Think Big, THINK BIG, American Mission, Jobs and Democracy PAC, JOBS AND DEMOCRACY PAC, Defending Our Values PAC and DEFENDING OUR VALUES PAC. 

In total, there were 336 advertisements in the dataset: 86 from Think Big, 57 from American Mission, 22 from Defending Our Values and 171 from Jobs and Democracy. Each ad listed on the platform was counted individually, although many of these were duplicates of the same ad that was run over a different period or platform. Some of the ads have also been removed for ad policy violations, but we were able to see which candidates they were for because they appeared when searching for ads specific to those candidates within the advertiser’s page. 

For the coding of themes, we looked at whether each of the themes mentioned (AI, data centres, economy, Trump, and immigration or law enforcement) appeared in any of the ads for each candidate. This does not account for how many times each theme may have appeared in ads for a candidate.

Our notes on the Google ad data and count of themes for each candidate can be found here. 

For the accompanying map showing spending by each Super PAC by state, we used data from the Google Ad Transparency Center showing the total spend per advertiser and the spending range per ad to estimate the spend per ad (as the data does not show the exact spend per ad). For each advertiser, we separately added up the minimum and maximum spend amounts for all ads. We then calculated where in that range the total spend fell (e.g. 45 percent between the minimum and maximum spend). The per ad spend was estimated by applying this same proportion to the range of each ad and rounding down to the nearest dollar.

FEC Campaign Finance Data

Data from FEC filings for LTF, Think Big and American Mission are based on disclosures from Aug. 20, 2025 to Oct. 14, 2026. Data from FEC filings for Public First, Defending Our Values and Jobs and Democracy are based on disclosures for Jan. 1, 2026 to Sept. 30, 2026, which were the latest available as of publication.

The amounts LTF and PFA contributed to their associated Super PACs are based on the recipient Super PACs’ receipts for contributions from “Leading the Future”, “Public First”, “Public First Action” and “Public First Action, Inc”. 

Links to the source pages for these filings, as well as the data we exported for our analysis, are available here.

---

This investigation was supported by a grant from the Tarbell Center for AI Journalism. 

Data visualisations by Galen Reich. Melissa Zhu contributed research to this piece.

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 Newsletter and follow us on Bluesky here, Instagram here, Reddit here and YouTube here.

The post Inside the Secretive, Contradictory World of Campaign Ads Funded by AI Super PACs appeared first on bellingcat.

Sign up here to receive Bellingcat’s biggest investigations by email as soon as they are published.

In September, the AI company Anthropic published a blog post that claimed its models were approaching “superhuman capabilities for geolocating outdoor photos”. Anthropic made this claim by comparing its models’ performance to that of players of the popular online game GeoGuessr, which challenges users to geolocate Google Street View images. 

Bellingcat has been running its own benchmark tests on AI models since June 2025, right around the time when Large Language Models (LLMs) began to challenge reverse image searches for assisting with geolocation tasks. In our first round, GPT 4o-mini from OpenAI appeared to be the best tool for geolocating, a title which was taken by Google AI mode when we performed our second round of tests a few months later in August 2025. 

This time, we significantly upped the ante for the models, providing increasingly challenging photos and asking for only one set of coordinates. We tasked not only closed-weight models from Claude, Grok, Gemini, ChatGPT, and Mistral, but also open-weight models from Moonshot AI’s Kimi, commercial provider GeoSeer and the non-generative geolocation model GeoClip. In total, we fed each model 52 photos that had not been published anywhere on the internet before.

The results showed that the current generation of models have markedly improved geolocation capabilities. Anthropic’s Opus and Fable, as well as Google’s recent Gemini 4 Argon model, performed the best on geolocation tasks, narrowing down locations from images faster than a human could under specific conditions. Preliminary tests of a set of agentic skills and tools being developed internally by Bellingcat also showed promising results in further improving the geolocation performance of generic-use AI agents.

But crucially, the tests also showed why humans remain necessary for accurate and trustworthy geolocation work, as no models were foolproof. This is particularly important given the increased use of AI in military contexts. The Anthropic blog post focused specifically on geolocating photos for military and intelligence purposes such as targeting, a historically labour-intensive process requiring significant human expertise. 

While our tests showed some models performed extremely well at narrowing down results to the correct general area, none were close to being perfect at finding the exact spot a picture was taken. That accuracy is extremely important in geolocation, especially in military contexts where the cost of incorrect information can be life or death. It is also important for organisations like Bellingcat whose reputation relies on pinpoint factual accuracy.

Our Methods

To compare model results, we used two metrics built on the distance in kilometres between a prediction and the real location. GeoGuessr Score was used to assess whether models could systematically identify the general area, while the Mean Log Score (explained below) measured which models most accurately pinpointed the correct location in the final few kilometres. 

The GeoGuessr Score metric is based on the online game of the same name, whose scoring system Stanford University academics reverse-engineered in a 2024 study. This system rewards answers close to the general area an image was taken, but does not strongly distinguish between guesses close to the precise location. For example, an answer 200km from the correct location with more points than one 2,000km away. However, an answer 500m from where a picture was taken and another 5km away achieve a similar score. For real-world geolocations, that distance matters a great deal.

For more precision, we introduced a custom metric that grades models higher if they narrow down the search in the last few kilometres. This metric, the Mean Log Score, assigns a value from 0 (completely inaccurate) to 100 (perfect model) and works on a logarithmic scale instead of a linear one like GeoGuessr score.

Therefore, a 10-fold improvement in distance means the same whether it is from 5km to 500m or from 500km to 50km. You can compare both metrics here.

Below are the model results sorted by either metric, showing how many photos each model geolocated within distance bands from the true location.

Several models placed at least 10 out of the 52 photos correctly within one kilometre, and the top-performing models placed over 50 percent of the photos within 25 kilometres.

NOTE: Geoseer-fast refused to answer on four of the 52 photos, claiming “This image is not suitable for geolocation”, while Mistral-large-4 refused once with “insufficient information in image to identify location”. Kimi-k2.6 also failed to provide a set of coordinates on one photo, and Gemini 4 Argon timed out repeatedly on one photo.

Our Test

To independently assess the models, we sought to give them very specific instructions and some images that were harder to geolocate than those provided in our past tests. The prompt asked for a single set of coordinates reflecting the photographer’s location. The purpose of restricting the response in this way was to prevent hedging, such as multiple guesses or confidence scores, and make the grading more uniform across different models. Models were allowed to use their providers’ built-in tools: search, code execution and page fetch. We capped their token output at 64,000 and ran each model under its default effort setting with a maximum time of 15 minutes per photo. If a model did not return coordinates on the first prompt, we also allowed a single rescue prompt asking it to do so using its previous reasoning and without using any tools.

The images provided were unique, from around the world, not published on the internet before, scrubbed of metadata and provided without any context. They did not contain significant landmarks. We also included images in snowy, foggy and nighttime conditions.

The results show that the LLMs tested could, in many circumstances, successfully narrow down a location much faster than a human can geolocate even without any context. This is not limited to Google Street View-style images. Take, for example, this image of vegetation and a very small part of a rooftop.

For an average person, identifying the country this photo was taken in without any context could take a significant amount of time. However, multiple models quickly narrowed it down to Kenya. You can navigate their answers in the map below. Notice how several correctly inferred enough vegetation details to narrow down the exact part of the world and the park itself, but also that for many, the final response was a best guess only. It is important to note that even though some models were correct, others were significantly off.

The full model results for all 52 photos can be seen in the table below.

Of the models tested, Gemini 4 Argon scored highest on the Mean Log Score, and Claude’s models (Opus and Fable) were at the top of the table regardless of the score metric. Even Claude Sonnet, the weakest of its lineup, consistently came close to the real location while being the fastest to answer across all models.

This was followed by GPT-6 Astra and GPT-6.1 Sol, as well as other Claude and Gemini models, and then the geolocation-specific model, GeoSeer Fast, which provided good results but also timed out frequently. We tested GeoSeer’s more competent Agent model, but it refused to answer on eight photos and regularly timed out.

For secondary factors like time and cost, Claude also performed well on both metrics, and Gemini Argon, while much slower, came at a similar cost to Opus. Based on API rates, Argon and Opus models offered the best balance of accuracy and cost, especially compared to Fable and Grok. Grok was the most expensive option and had an underwhelming bottom-half performance. Meanwhile, GeoClip was freely available but only outperformed Claude Haiku.

Model Limitations

While we sought to analyse the most prominent models, there was one that we could not include as it has yet to be fully released.

In its blog post, Anthropic claimed its Mythos model beats “even the strongest human baseline on median distance error, scoring 37.0km and 47.2km across 6,000 photos (placing 23.7 percent and 23.1 percent within 1km).“ But Nuno Moniz, Associate Research Professor at the Lucy Family Institute for Data & Society at the University of Notre Dame, told Bellingcat that “the median represents a strong simplification of what would be acceptable to compare AI-based system performance with regards to humans”. 

Anthropic compared its models to competitive GeoGuessr players and their success in identifying the location of Google Street View images. The entire set of test images was online and there was no guarantee the models were processing the imagery for the first time. 

While Anthropic has been relatively open about these methods, not all developers have been as forthcoming. It is therefore difficult to assess whether the same limitations will impact all other models, or whether there will be further caveats that we are simply not aware of. 

Moniz said that “the increasing availability of benchmarking tools and datasets has made it somewhat difficult to measure reliability in test results such as these”. Specifically, due to the ”uncertainty as to whether that data was included in its training”. This is why we ran our Bellingcat tests with non-published photos sourced from Bellingcat staff and the Bellingcat Volunteer Community.

A further limitation of any models comparing performance to GeoGuessr scores is that the game uses images from Google Street View, which is limited to a select number of countries, and within these, a limited number of roads. Street View is also intentionally skewed to feature clear, high-quality photos taken during daytime that are useful for navigation, and users can pan around their view. There are also some well-known tricks that allow GeoGuessr players to bypass actual geolocation to correctly guess the location of photos: for example, Google is known to use a black car in Jordan and a white car in the UAE to capture imagery. While this wouldn’t necessarily impact model performance, it is a drawback of measuring against GeoGuessr score as the game has a set of workarounds that can bypass geolocation skill.

Anthropic’s data consisted of thousands of social media photos of various scenes. Once again, though, these photos were online and may have featured in the model’s training data. Anthropic’s blog post contained four images as an example of this dataset.

The examples Anthropic shared are relatively easy to geolocate as they feature recognisable buildings and signs. But footage that requires geolocation often comes from regions where access is limited and which do not feature on social media or Street View.

Specifically, in a military and intelligence context, footage often involves low-resolution material and intentionally obscured details recorded in remote or inaccessible areas. Context that could be found online can be uncertain or plain wrong, especially during a developing situation, such as the image below that Bellingcat investigators geolocated when verifying a US strike on a school in the city of Minab in Iran that killed at least 120 children earlier this year.

Anthropic researchers acknowledge this caveat, writing that it “does not as clearly emulate the task of precisely pinning down a location in near-real-time in a less populated battlefield setting; that is a task for future research.”

As noted, it is also possible that similar caveats need to be applied to other models when seeking to understand and interpret any geolocation-related benchmarking.

Supercharged Models

In parallel to the above models, we also tested a set of agentic skills and tools being developed internally by Bellingcat. These are researcher-informed instructions and code that can equip generic-use AI agents to augment their geolocation ability. For instance, we designed skills such as allowing it to download Environmental Systems Research Institute (ESRI) satellite imagery tiles. We also designed another skill that enabled direct access to data from Bellingcat’s OpenStreetMaps search tool, which aims to assist geolocation efforts by comparing items identified in imagery (like the position of a pylon compared to a nearby railway line) to OpenStreetMaps data. GeoSleuth is a similar project that has already been published, which we used for comparison. We picked Claude’s Opus 5.5 model to run both skills customisations because of its strong performance and reasonable pricing, and ran all the photos through the supercharged models under the same conditions – except that they had no timeout enforced as the supercharged models required more time to utilise their extra skills. 

The different timeout conditions and expanded toolset do not allow for a direct comparison of the customised models to the previous off-the-shelf models’ results, even if they both outperformed the barebones Opus model they were based on, and all other models on GeoGuessr Score, with only Argon ranking above Bellingcat’s agent in Mean Log Score.

Still, this test allows us to showcase why skill customisations can bring value. Take the following photo.

This is not an exceptionally hard geolocation, and 11 of 17 models answered within 5km of the actual capture location in Venice, Italy. GPT-Sol answered with the correct monastery, but only Bellingcat’s agent and GeoSleuth answered with the exact camera position coordinates. This was possible because both model customisations can download and compare satellite imagery tiles, which allows them to validate not just which monastery the photo was taken at but also the correct courtyard position and orientation from visual cues. You can click the buttons in the list below to audit each model’s final answer.

While in this instance the supercharged models got the exact spot, there were other instances where they did not, as the scores in the results table above show. This again emphasises that even with extra skill sets added, these models were still far from infallible.

Tests Conclusions

Despite increased model improvement, our analysis shows that important caveats remain and it is important to keep a human in the loop when seeking to guarantee a geolocation is precise and correct.

Our tests show that the coordinates models provide as answers are rarely an exact match, even with supercharged models. Invariably, even the best models, with the best tools, can confidently claim a wrong position. Anthropic says its Mythos model was able to pinpoint less than a quarter of its social media photos dataset within a 1km radius. But narrowing down a location is rarely the trickiest part of geolocation work. A large part of important geolocation work is in the final stretch: finding specific coordinates from a general area. This level of specificity is needed because a general area is not a match. Matching features and perspective are necessary for an accurate geolocation, and in many cases that is done on very low-quality images or even by cross-referencing multiple sources. 

This level of precision is essential for investigations where it matters what building a person could be found at, what lines of sight individuals might have had, or what impact an air strike might have led to. 

Since LLMs can still confidently get things wrong, as our analysis shows, all results still have to be verified independently. Strong human geolocation skills therefore remain crucial, as the more one understands about geolocations, the better one can verify or take to completion an automated result. 

Though it can speed up geolocation work significantly, especially in narrowing down an area of interest, AI models remain a useful assistant rather than a tool we can blindly delegate geolocations to.

---

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 Newsletter and follow us on Bluesky here, Instagram here, Reddit here and YouTube here.

The post Can LLMs Supercharge Geolocations? Our Tests Show Big Improvements But Humans Remain Key appeared first on bellingcat.

Between February 2022 and September 2025, Bellingcat staff and volunteers collected, geolocated, and shared more than 2,500 incidents of civilian harm following Russia’s full-scale invasion of Ukraine. 

As part of this effort, Bellingcat tested a new machine learning model intended to rank Telegram social media posts on their likelihood of containing incidents of civilian harm. 

This novel methodology dramatically reduced the search and selection time required, freeing researchers to focus on verifying incidents of civilian harm – not just searching for them. 

This piece documents our methodology, ethical considerations and lessons learned in the hope that others researching similar topics can benefit from our work. 

Open source research into civilian harm is still a relatively new field and it presents many challenges – one of the biggest is organising and sorting through the huge volume of user generated content being produced to find what is relevant. 

Machine learning, a form of artificial intelligence that uses algorithms to identify patterns from large amounts of data and make predictions, can make this task more efficient.

With ongoing conflicts involving large amounts of civilian harm occurring in Sudan, and much of the Middle East, this guide aims to offer those covering these conflicts an example of how machine learning can be used to help find and sort incidents. You can also access the Code Notebook for our model here.

We defined “civilian harm” not just as civilian deaths or injuries resulting from armed conflict, but also the broader and delayed effects on civilians from mental trauma, loss of livelihood, displacement, destruction of infrastructure and more. This definition was informed by the Protection of Civilians book on civilian harm. 

Initial Telegram Dataset 

Each Telegram post containing civilian harm which had already been manually verified by researchers was used to build an initial dataset of confirmed cases of civilian harm, which data scientists call positive instances. We collected a total of 5,848 unique URLs for these Telegram posts. For our manual collection we reviewed posts on relevant Telegram channels, working through oldest to newest posts each day. Assuming that a given post made it to our geolocated incidents list, it meant the researcher who flagged it also looked at the posts that appeared before and after it on Telegram and did not flag those ones, so we selected the 10 posts surrounding the verified civilian harm post as our additional dataset of posts that did not contain civilian harm. After excluding any deleted or duplicate posts, we ended up with 48,545 non-civilian harm posts, our negative instances. 

The choice to overrepresent negative instances aims at better reflecting the real world and increasing data available for model training. 

We enriched each URL with metadata from the Telegram API, such as the time of publication, reactions or textual content. As some of these posts had been deleted, we completed the missing data points with previously preserved versions from our Auto Archiver database, only available for the positive instances.

Feature Engineering

Training a machine learning model requires numerical data, as these models compute a prediction score based on mathematical operations.

We built these by converting raw data from our initial dataset, such as keywords signalling potential civilian harm, into numerical scores (or “features”) that the model could interpret, with the aim of increasing the model’s ability to identify patterns. This process, known as feature engineering, can significantly improve model results because it allows data scientists to suggest explicit context knowledge. 

A full list of features we used to train the model can be found in the code notebook accompanying this piece. Many features were directly inspired by researchers’ input from their experiences manually screening cases of civilian harm by sorting through a set number of Telegram channels and inspecting each post individually.

Several of the features used were directly built from the metadata contained in each Telegram post including media_type, day_of_week; or binary ones: forwarded, edited and reply_to. 

Other features included engagement information: views, forwards, total_reactions, and even individual features for most used emojis including the reaction_crying_face to count emoji.

Converting Text to Numbers 

To embed the experience from the manual collection process, researchers put together a list of keywords both in Ukrainian and Russian that, to them, signalled posts likely to  show civilian harm. For instance, “Шахед” and “КАБ” translated to “Shahed” and “Guided aerial bomb” respectively. We created a numerical feature to count their frequency. 

In addition, we included several generic English-language keywords which meaningfully signalled potential civilian harm, such as “injured”, “school affected” and “hospital affected” that were only used for generating semantic similarity scores. 

A semantic similarity score is a calculation used to determine the proximity in meaning between different words and phrases. To get the semantic similarity between the post text and each of our keywords, we represented each in a list of numbers via a Sentence Transformer model, which converts words into numerical representations called vectors that a computer can understand. 

We then calculated the level of similarity between each vector using cosine similarity, one of the most popular methods for measuring similarity between two pieces of text.

Due to how embeddings work, this calculation results in a figure on a scale from -1 (no semantic proximity) to 1 (same meaning). For example, the words “hurt” and “injured” would have a high similarity score, while “residential” and “injured” would have a negative score as the words are not semantically similar. 

Finally, to enable the model to identify the relevance of each post to civilian harm in Ukraine, we used a multilingual text transformer from the BERT family of language models to represent the entire post’s text as a vector of 768 numerical values. This model can efficiently represent text from many languages in a way that captures meaning: the same sentence in different languages will generate similar embeddings, and trained machine learning models can detect patterns in the embeddings. 

It is important to note that for this initial prototype of a civilian harm detection model, we did not include any features derived from media content such as photos and videos, although that would be a logical next step in attempting to improve model performance.

Selecting, Training and Evaluating Models

With 54,393 rows of 893 numerical features each, we selected four machine learning algorithms to train our predictive models. 

We chose Logistic Regression as a baseline algorithm due to its simplicity. We also selected three other “best in class” models, Random Forest, XGBoost, and LightGBM. These choices centred on the interpretability of the models and their ability to work on tabular data of this size. For example, we avoided neural networks due to a lack of interpretability and because those models work best with a larger dataset. 

To genuinely assess the performance of the trained models, we split our dataset into three parts:  

• A training set – the data the models were trained on (60 percent of the full dataset’s rows)
• A validation set – used for an intermediary evaluation when tuning model parameters (20 percent of all rows)
• A test set – hidden for the final performance assessment, so the models were evaluated on unseen data (remaining 20 percent of rows)

We used a stratified split to divide the dataset instead of a random split. This method ensured the proportion of positive instances (i.e. confirmed cases of civilian harm) remained consistent across all three sets at about 11 percent.

To measure the performance of machine learning models, we ran them through the test set and measured the number of correct and incorrect predictions. Models output a likelihood between 0 and 1 that each Telegram post contains civilian harm, and we tried to find a cut-off threshold that leads to a good balance between flagging almost every post (0.1) or flagging very few (0.9). 

There are two main types of evaluation metrics to gauge a model’s prediction power. Recall asserts what fraction of positive instances (i.e. known civilian harm posts) were correctly flagged as such. Precision measures the fraction of posts flagged as civilian harm that are indeed civilian harm posts.

During the training phase, we tuned the models to maximise average precision (PR-AUC), a metric that summarises precision across all recall levels. While this method also accounts for precision, it prioritises recall, which is preferable for this use case as it steers model selection to reduce the number of civilian harm posts that are skipped. 

The following table sorts models from best to worst PR-AUC against a baseline of a coin-flip predictor. ROC-AUC and F1 are two other evaluation metrics included as sanity checks. Simply put, ROC-AUC measures the probability of ranking two instances, one negative and one positive, correctly; F1 balances precision and recall equally and its best cut-off threshold value.

From these results, we selected XGBoost as our final model as it had the best scores when compared across all metrics.

Interpreting the Model

Because these models are interpretable, we can understand which features are the most useful when predicting whether a post includes civilian harm. The above table shows the top 10 features that most strongly signal the XGBoost model to make a decision:

• semantic_keywords_similarity: the semantic proximity between the post text and manually selected keywords “casualties”, “damage” and “civilian harm”
• bert:  the model was able to discern meaning from the text with the same strength as some of the other features in this list – there are three cases of this in the top 10
• reaction_crying_face: reactions with crying face emojis on the post
• group_of_messages: whether a post contains multiple media files
• keywords_in_text: the number of custom Ukrainian or Russian keywords in the post

These results generally tally with what you might expect when selecting Telegram posts for instances of civilian harm, including that posts that generate a lot of emotional engagement and posts using keywords about civilian harm were among those most likely to contain content related to this topic. Not all models had the same top features as XGBoost. In fact, for the Random Forest model the most important feature was the number of crying face emojis present in a post, a soft pattern highlighted by researchers when this methodology was first imagined.

LLM Results and Comparison

Retroactively, we decided to run a sample of the same test dataset through different large language models (LLMs) to gauge their ability to make these same predictions. 

We aimed to include an LLM-generated score as an extra feature for our trained models, which would be captured as relevant if it correlated with the correct predictions. 

To start, we selected two local models, the 1B and 4B variants of Gemma 3 from Google DeepMind, and two cloud-hosted models, Gemini 2.5 flash and Gemini 3.5 flash. With this selection, we hoped to compare results across a wide range of models’ expected performance. 

We generated a 400-row stratified sample (preserving the same proportion of real civilian harm instances) from the test dataset used for the custom models. For each of the four LLM models, we ran two tests: one where only the Telegram post message was sent, and another including both the message and the engineered features (excluding the text embeddings, as the model had direct access to the text). In the prompt for each model, we asked for a score between 0 and 1. We then evaluated the results as we did for the custom models. 

The above table shows that LLMs can indeed extract value from the engineered features. All four LLMs surpassed the baseline Logistic Regression model in our tests, yet none of them performed better than the other custom-trained models, and XGBoost remained the one with the highest PR-AUC. 

Still, Gemini 2.5 Flash performed better than its newer version 3.5 and even achieved a slightly higher best F1 score than any other model. While this is a good result, for the flagging of civilian harm posts, the PR-AUC remains the crucial metric, as it captures the model’s ability to identify infrequent instances of civilian harm while minimising false positives.

Ethical Considerations

Introducing an instrument of automated decision-making into a process of detecting civilian harm brings inherent ethical questions. These include automation bias, or how humans tend to blindly place faith in machine-generated recommendations; algorithmic bias, or how the results of these models echo the same patterns present in the training data, including under- or over-representation of types of civilian harm. 

The decision to test an automated methodology for this particular project came from the fact that there were limited resources for both steps in the process – the detection of potential civilian harm and its actual verification. Historically, we built an enormous backlog of unverified incidents because a lot of time had to be spent on monitoring the most recent events so that potential evidence would be captured and preserved as soon as possible. 

The automation of this process also reduced the exposure of researchers to a significant amount of unpleasant and distressing visual and text content, reducing the burden of exposure to traumatic content. 

For this project, we tried to ameliorate the ethical challenges with a number of strategies including randomly flagging posts not captured by any model, monitoring which features models relied on to make decisions, and by doing historical comparisons of patterns in data. 

Additionally, as stated above, for this initial prototype of a civilian harm detection model we did not include any features derived from the media content itself. In the future, it would be a logical next step in attempting to improve the model performance, to include the media from the posts – but using AI to review actual media comes with additional ethical challenges such as model bias.

Because of the opaque ownership of many LLM companies and their generative nature, the use of LLMs for an extra feature presented additional ethical challenges including privacy and safety concerns considering the sensitive nature of the data. Our model did not rely on LLMs, though we retroactively ran a sample through it. 

How the Model Fits into the Bigger Picture 

After selecting this model, we created a user interface where researchers could view a list of Telegram posts sorted from most to least likely to contain indications of civilian harm. The user interface was designed for quick triage and integration, where a positive confirmation from researchers would instantly send the post to the Auto Archiver (Bellingcat’s tool for preserving digital content) and then transfer it to ATLOS (our internal collaborative verification platform). Bellingcat staff and volunteers could then manually verify incidents. Researcher input was constantly stored so that this data could be used to improve the model in the future. 

Preliminary feedback indicated that the AI model was useful. Not only were we able to reduce time and harm from scouring through dozens of war reporting Telegram channels, researchers also reported that the stream of new posts being added to the verification backlog were capturing real and diverse cases of civilian harm. 

We recognise this model has much room for improvement and is a work in progress. Even though it can illicit diverse civilian harm posts, further tests and improvements (such as improved feature engineering and continuous evaluation) are needed before it can confidently be deployed.

Despite the focus on civilian harm and Telegram (highly popular in Ukraine and Russia), this pipeline is generic and can be adapted to other conflict monitoring tasks. How easily this can be done does depend on how open the social media platform is and whether it is possible to scrape posts from it. Apart from that, it is easy to incorporate new features and data, and cheap to automatically retrain, test and deploy models as the system receives more human input.  

Looking forward, sorting through overwhelming amounts of data in a conflict will continue to be challenging. Hopefully, this methodology can help newsrooms, conflict monitoring organisations, and others find the balance between ethical considerations and resources in order to carry out open source investigations on civilian harm and human rights violations. 

---

Editor’s note: This article was updated on July 3, 2026, to include a line outlining that the model described is a work in progress.

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 How to Use AI to Help Find Civilian Harm appeared first on bellingcat.

The video posted by a state branch of India’s ruling Bharatiya Janata Party (BJP) showed Assam chief minister Himanta Biswa Sarma shooting an image of two men in Muslim skull caps. “Foreigner-free Assam”, read one caption across the video. “Why did you not go to Pakistan?” said another. 

One of the men in the photo that Sarma was portrayed as shooting was Gaurav Gogoi, a leader of the Indian National Congress (INC), the BJP’s main competitor in Assam for the state’s upcoming legislative elections next month. 

Gogoi has stated that he is Hindu but enjoys visiting different religious sites and observing their norms. He has been photographed wearing traditional Muslim attire during religious occasions such as Eid. 

But the image of him in the video shared by BJP Assam, wearing a casual singlet with a skull cap, was not one of those occasions. 

Bellingcat has seen several dozen videos posted by the BJP that use generative artificial intelligence (AI) alongside anti-Muslim and anti-Bangladeshi messaging in the border states of Assam and West Bengal in December last year, ahead of legislative elections scheduled in both states for April.

Bellingcat analysed 499 social media posts containing photos and videos shared on Facebook, Instagram and X by the BJP’s official accounts in the two states for this time period, finding 194 posts that appeared to meet the United Nations’ definition of hate speech: discriminating against persons or communities based on inherent characteristics such as religion and national origin. Of these, 31 (about one in six of the hateful posts) contained the obvious use of AI-generated imagery. 

Chart: Galen Reich

These appear to be part of a larger pattern of politicians and parties globally using generative AI to amplify hateful or divisive content, particularly ahead of major political events such as elections. 

Ahead of the New York City mayoral race last year, Andrew Cuomo’s official X account shared, then deleted, an AI-generated video depicting Mamdani eating rice with his hands and a Black man in a keffiyeh shoplifting. In Italy, several opposition parties complained to a communications watchdog after deputy prime minister Matteo Salvini’s League party published a series of AI-generated images depicting men of colour attacking women or police officers. And in the UK, videos by an AI-generated rapper funded by the far-right Advance UK party, with lyrics targeting Muslims, were viewed millions of times. 

A Campaign of Hate

Both Assam and West Bengal share a border with Bangladesh. BJP, the world’s largest political party, is currently in power in Assam, where legislative elections are scheduled on Apr. 9. West Bengal, which goes to the polls on Apr. 23, is governed by the Trinamool Congress (TMC).

Tensions between India and Bangladesh worsened after former Bangladeshi Prime Minister Sheikh Hasina, who enjoys close ties with Delhi, was ousted in 2024 and fled to India. 

US-based international affairs expert Mohammed Zeeshan told Bellingcat that the “dehumanising and debasing” terminology used in India to refer to alleged illegal Bangladeshi immigrants, including by senior ministers, has caused resentment towards India in Bangladesh. 

“The situation, in fact, was so bad that Hasina herself had subtly warned the Modi government in public statements that Indian domestic rhetoric was endangering Bangladeshi Hindus, who bore the brunt of that resentment,” Zeeshan said. 

Zobaida Nasreen, a professor of anthropology at Dhaka University, said that anti-Muslim rhetoric intensified by BJP leaders reinforces the belief in Bangladesh that Muslims and Bengalis are being collectively targeted in India.

“Viral videos containing this message tend to spread quickly across Bangladeshi media and social platforms especially on Facebook, enhancing perceptions of hostility and triggering anti-India sentiment or nationalist backlash,” she added.

In December, the month our dataset was collected, Dipu Das, a Hindu garment worker, was beaten to death at an anti-India protest in Bangladesh over allegations that he had made derogatory remarks about Islam. 

And while the administration led by Bangladesh’s newly elected leader Tarique Rahman has sought to reset strained ties, most of the hateful social media posts we saw posted by the BJP in December attacked Bangladeshi Muslims and/or Bengali-origin Muslims in India, showing how tensions between the two countries continue to influence political messaging in India’s border states.

Bellingcat’s analysis included a total of 202 posts by BJP Assam and 297 by BJP’s West Bengal branch on their official accounts. We also looked at posts shared by BJP’s main opponent parties – 194 from INC in Assam and 357 from the TMC in West Bengal – during the same time period in December. 

This included all visual social media posts (containing photos or videos) by each party in December, except those that did not appear to contain any overt political messaging, such as those simply commemorating public holidays. We only counted each photo or video once, regardless of how many platforms it was shared across. 

Although all of the major parties contesting in the Assam and West Bengal state elections appeared to use AI-generated imagery in some of their posts, there appeared to be a particularly high concentration of hateful messaging in the ones posted by the BJP’s accounts. 

In Assam, we identified 28 posts by BJP using apparently AI-generated imagery, of which 24 carried hateful messaging. Of the 194 INC posts we looked at from December, 41 appeared to feature AI-generated imagery, but none of these appeared to carry hateful messaging. 

In West Bengal, we found 14 BJP posts that contained clear indicators of AI-generated imagery, seven of which were hateful. We also identified 15 posts by the incumbent TMC that appeared to feature AI imagery, but none of these appeared to meet the definition of hate speech. 

When contacted for comment, BJP Assam spokesperson Rupam Goswami did not directly respond to questions on the party’s general use of AI but said they did not post any AI-generated photos of Gogoi. “BJP does not stoop so low,” he told Bellingcat.

As for the “point blank” shooting video, Goswami initially said the person responsible had been punished and removed from the party. However, when asked about Sarma saying that he would re-post the video with those he was depicted shooting labelled as “Bangladeshis”, Goswami said, “[Bangladeshis] need to be completely suppressed.”

BJP West Bengal did not respond to multiple requests for comment by Bellingcat via phone and email.

It is important to note that as generative AI technology improves, it can be increasingly difficult to detect AI-generated imagery. Our manual count of AI-generated imagery only included posts that had obvious signs of generative AI such as unnaturally smooth textures and multiple people with the same faces. It is therefore possible that there were other images in our dataset where generative AI was used more subtly. 

However, Joyojeet Pal, Professor of Information at the University of Michigan, told Bellingcat that the quality of these visuals, or whether they looked real, was not the priority. 

“What politicians in India have understood is that the sociocultural drivers of misinformation are most important for elections, so they harp on about things to the extent that they have started to not care about form over substance. It looks bad? It doesn’t matter,” he said.

More important to voters, according to Pal, was whether they already believed in the narrative contained in the videos, which generative AI could help create more quickly: “AI is helping cement polarised opinions by giving you the kind of content you have already decided you want to engage with.” 

When asked about INC’s use of AI, party spokesperson Aman Wadud said that it was obvious that some of the videos they posted were made with AI and that there was no intention to mislead. 

“AI can be both destructive and creative. We are using it in a creative manner, we are not using it in a destructive manner. We don’t violate people’s dignity, we don’t falsely accuse people,” he said.

TMC did not respond to Bellingcat’s multiple requests for comment via phone and email by publication time.

Portraying Bengali Muslims as ‘Foreigners’

The largest category of hateful messaging Bellingcat observed in the BJP’s posts targeted Bangladeshi or Bengali-origin Muslims, referring to them as “infiltrators” or “foreigners”. We counted 66 such posts by the BJP’s Assam and West Bengal branches from December, of which eight appeared to contain obvious AI-generated imagery. 

Bengali-origin Muslims are often stereotyped as “illegal immigrants” in the state, although members of the community have lived in India since the late 1800s. 

Last year, the BJP deported thousands of alleged undocumented migrants – reportedly including Indian Muslim citizens – to Bangladesh. Human rights groups have called the deportations unlawful and discriminatory, as well as lacking in due process. 

One video referencing this theme shows AI-generated visuals of protests against “illegal infiltration” in Assam, with the caption urging people to “wake up” or the country would “turn into Bangladesh”. 

A different one uses real footage from past violence in Assam mixed in with images of Muslim men. A song playing in the background accuses them of taking over “Assamese land” and shows AI images of “Assamese” people, i.e. those not in stereotypical Muslim clothing, crying.

Both videos use religious markers to draw a distinction between “infiltrators” – men in skull caps or lungis associated with Bengal-origin Muslims – and “citizens” in non-Muslim attire. 

Clothing is often used by the Hindu far-right as a visual shorthand for identity and a deepening religious divide. In 2019, Prime Minister Narendra Modi said of protests against a controversial citizenship law that those responsible for violence could be “identified by their clothes”. 

In the hateful posts seen by Bellingcat, both real and AI-generated images of opposition figures – particularly Gogoi – were shown alongside messaging that suggested that they supported “foreigners” or “infiltrators”. 

The Center for the Study of Organized Hate (CSOH) also noted, in a 2025 report on AI-generated imagery and Islomophobia in India, that Hindu far-right politicians and media outlets have invoked and reinforced the trope of Muslims as “infiltrators” for years. 

“AI-generated images on these themes reinforce associations between Muslim identity and illegality, reinforcing xenophobic and Islamophobic stereotypes. In doing so, they play a powerful role in justifying exclusionary policies and normalising discrimination against Muslims,” the report said. 

‘Save Hindus’

Zenith Khan, a data analyst who worked on the CSOH report, noted that AI-generated propaganda was often tightly knit with current political moments, and its impact depended on “timing it right” especially when “people are emotionally charged”. 

The violence against the minority Hindu community in Bangladesh has been used by the BJP to raise concerns over the safety of Hindus in India. 

Days after Das’ lynching, the Assam state branch of BJP posted a video with an image of his face – except that it was manipulated with AI to show tears streaming from his eyes. “Save Hindus”, said the text accompanying the video. 

Posts by BJP’s West Bengal unit also seemed to frame Muslims as criminals or threats. A video, styled after the TV show “Stranger Things”, raised alarms over an “upside down” version of the state under the current government. 

A man is depicted being chased by men in skull caps. Arrows label them as “Ralib,” “Galib,” and “Chalib” – a play on Muslim names ending in “-lib” – in case the skull caps left any ambiguity about their Muslim portrayal. 

INC filed a police complaint in September last year against the BJP for sharing AI videos targeting Gogoi and the Muslim community, as well as another complaint in relation to the video of Sarma portrayed as shooting two men “point blank” in February. 

INC Assam spokesperson Wadud said that no action had been taken on the party’s police complaints as far as he knew. 

Disinformation researcher Bharat Nayak told Bellingcat that it has always been tech platforms’ responsibility to control new types of content. 

“The goal post can’t shift. This has always been a tech problem,” he said. 

When this responsibility is shrugged off, Nayak added, the result is a lack of accountability. “If you’re using old videos from other countries as new, you will have people countering you. But AI-generated videos can be shared without context just to spread hate – like showing people in skull caps – and the ‘when, where, how’ questions vanish.”

Both Meta – which owns Facebook and Instagram – and X have policies against hateful conduct. 

Meta also announced in 2024 that it would start adding “AI info” labels to more content detected as AI-generated, while some X users spotted a similar feature introduced on the platform last month. Only five of INC’s AI visuals that we identified – and none of those by TMC or the BJP – had a disclaimer that said “AI-generated”. 

Bellingcat reached out to Meta and X for comment on whether the posts we identified breached their terms of use regarding hateful conduct or labelling AI-generated posts. A Meta spokesperson said they were reviewing the flagged content and “will take appropriate action on any violations of our policies”. As of publication, X had not responded.

---

Kalim Ahmed from Bellingcat’s Discord Community contributed research to this piece.

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 How India’s Ruling Party is Using AI to Boost Hate Speech in States Near Bangladesh appeared first on bellingcat.

The Netherlands’ largest newspaper, De Telegraaf, recently published an interview with a woman claiming to organise her own evacuation flights from Dubai, selling seats at €1,600 (US$ 1850) each. Four days later, her photo was removed from the article, though the interview remained.

Bellingcat has found that the original image not only includes artefacts commonly associated with generative AI, but that the flights referenced in the article do not appear to exist.

The story came at a time when thousands of Dutch people were reportedly seeking urgent ways to leave the region following Iranian missile and drone strikes across the Gulf in retaliation for US-Israeli strikes.

Published on De Telegraaf’s website on March 5, the headline reads: “Dutch people in the Middle East feel abandoned by the government: We just rented a plane ourselves.”

The Dutch minister of foreign affairs was confronted with this headline during a television interview, in which he described ongoing efforts by the Dutch government to repatriate citizens to the Netherlands.

The article features interviews with several Dutch people struggling to leave Dubai and Abu Dhabi, including Tamara Harema. Under the subheading “Dutch people hire their own plane”, Harema says she was “rebooked five times by Emirates” and that the official repatriation flights organised by the Dutch government were not ‘taking off’.

As part of a group, she says, they are organising buses and have hired an Airbus A321 to fly home. Harema is quoted as saying: “The first plane is already full, so we’re organising a second flight. Stranded travellers can contact us.”

However, several discrepancies in Harema’s photo, published in the original article, suggest it was AI-generated. No trace of a person matching Harema’s face or profile could be found, and flight-tracking data suggests no such plane took off.

The Photo

In the image below, the world’s tallest structure, Burj Khalifa, can be seen through the window overlooking the Dubai skyline. Each side of the tower is unique, with platforms that protrude at different heights and in different directions. It also contains several mechanical floors, which appear as dark bands in the photo.

By cross-checking the height of the visible platforms together with the location of the mechanical floors, it’s possible to determine that Harema’s hotel room faces north-west, towards the Burj Khalifa’s south-east-facing facade.

Several discrepancies are visible when comparing Harema’s photo with other images of the building, including an upper mechanical floor appearing higher than in other images and the absence of the water feature at the base of the building.

To establish whether Harema’s photo could have been taken several years earlier, Google Street View imagery was analysed from 2013 onwards. No match could be found when comparing the arrangement of buildings at the base of the Burj Khalifa.

Several other irregularities, as shown below, including the hotel room furniture and details of Harema’s clothing and jewellery, also suggest it may have been AI-generated.

Fully Booked Airbus A321

Regarding whether the plane existed, Harema says in her interview that buses have already been arranged to collect passengers from two locations in Dubai on Saturday, March 7, after which a 232-seater Airbus A321 will depart from Muscat, Oman, for the Netherlands.

The article notes the cost is €1,600 (US$ 1850) per person, without detours. “Although we read that a Dutch repatriation flight costs €600, just try getting on such a flight,” says Harema.

According to Flightradar24, multiple A321s departed Muscat on March 7 and 8, but none bound for the Netherlands. The only aircraft that did arrive in Amsterdam from Muscat were either government-organised repatriation flights or scheduled Oman Air services, none of which were Airbus A321s.

Two Airbus A321s were recorded on the ground at Muscat Airport on March 7. One, belonging to Gulf Air, later departed for Rome via Riyadh March 8. The other, operated by SalamAir, had been flying routes between Oman and Bangladesh until March 3, but has since remained in Muscat.

After contacting De Telegraaf, an explanation for the photo’s removal was added at the bottom of the article, stating that the photo did “likely not meet our journalistic guidelines.”

The newspaper’s deputy editor-in-chief, Joost de Haas, added:

“Regarding the quoted Tamara Harema, the editors contacted her after Mr. Chizki Loonstein—a long-standing source for one of our reporters—informed us about attempts to charter a plane. Mr Loonstein informed us that Ms Harema stayed in Dubai and could tell us more about it. This led to messages from which several quotes from Harema were extracted, as reproduced in the relevant passage of the article.”

A search for Loonstein led to a six-month-old report from another Dutch newspaper, NRC, which claimed that Loonstein, a lawyer, emigrated to Dubai after his legal company went bankrupt, leaving his clients, victims of fraud, worse off.

Contacted for comment, Loonstein confirmed that he knew Harema and had shared her contact details in “an app group” in relation to a flight from Muscat to Amsterdam. After this contact, Bellingcat sent him the photo of Harema to confirm her identity and asked him to share Harema’s contact details. In response, Loonstein refused to provide further comment. 

---

Merel Zoet and Claire Press contributed to this report.

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 AI Used to Promote Non-Existent Evacuation Flights From the Middle East appeared first on bellingcat.