Freedom Index
![As news broke that 84,270 people, including more than 32,000 children, had been affected by Nepal’s catastrophic floods, one visual offered something the world desperately needed in the moment: hope.
The video shows a little girl, buried under thick mud, being is rescued by Nepal’s emergency teams. Instinctively, the viewer roots for her as rescuers carry her to safety, wrapping her shivering body in a red shawl. But the most moving moment of the video comes towards the end. The girl is sitting motionless in a relief camp, an officer feeding her warm soup. As the camera lingers on her face, a faint smile almost begins to show. It is the kind of image that can offer the viewer respite from the horrors of a humanitarian crisis. The video rapidly spread rapidly across all social media platforms, with commentators calling it a miracle.
Of sorts. In fact, it was an AI miracle.
Agence France-Presse (AFP) Fact Check’s investigation led them to an India-based creator, who admitted creating the video with the help of AI-generative tools. His justification was disarmingly straightforward: emotionally engaging visuals perform well online.
This was just one of the many digitally-manufactured “miracles” put into social media heavy rotation during the floods. Another clip, showing a bridge collapsing into floodwaters swallowing vehicles and human lives, seemed to have been lifted straight out of a Roland Emmerich disaster film. The video racked up more than 2 million views on X before it was debunked as AI-generated by Factly. What made this social media post stand out for me was the identity of one of the people sharing it – an experienced news editor. The post, still pinned on his X profile and with more than 2 million views, suggests that even experienced newsroom decision-makers could be fooled by synthetic disaster footage – or, worse still, be untroubled about the veracity of such videos.
Both videos incorporate the visual language of short films, but were not released in cinemas or on Netflix. They surfaced on Facebook, X and Instagram, even as emergency teams continued to search for the thousands of people that went missing after a glacial collapse triggered devastating floods in Nepal and Tibet on August 26. Even as survivors photographed authentic eyewitness footage, synthetic visuals emerged alongside them. The floods became one of the first major climate disasters in South Asia where documentation and fabrication coexisted in the same visual ecosystem.
Screenshot of debunked Nepal floods footage
“During earlier disasters, our primary challenge was identifying old footage, establishing where a video was recorded and determining whether it was being presented in the correct context,” Saurabh Shukla told Index. Shukla is founder and editor-in-chief of NewsMobile, an India-based independent media organisation that debunks AI-related deepfakes and focuses on verified information. “The biggest difference today is the sheer speed at which fabricated and misleading visuals entered the information ecosystem,” Shukla said. “With generative AI, we now have an additional challenge – determining whether the event depicted in a video ever happened in the first place.”
This has fed into a media environment where unverified visuals are disseminated faster than the verified facts reported by journalists. “Speed without verification can amplify confusion, while verification without timely reporting can leave an information vacuum,” Shukla said. “The challenge for newsrooms is to strike the right balance between being fast and being accurate.”
The first images of rescue missions in secluded mountainous regions almost always come from locals, and thus forming the spine of disaster reporting. These inevitably blurry and shaky recordings do not have the polished quality of professional news reporting, but do carry something far more consequential: the credibility that comes with eyewitness documentation. These images often shape the core of the collective memory of a tragedy. But in an age of visual manipulation and trickery, that presumption of authenticity no longer holds.
Factly conducted 30 fact-checks into coverage of the Nepal floods; fully one-third involved AI-generated visuals. “Beyond checking the location, source, and context of a visual, we also had to establish whether the visual itself was authentic,” Akshay Kumar Appani, lead fact checker and researcher at Factly, told Index. “During breaking disasters, old AI-generated visuals can spread rapidly while being presented as current footage. This makes it harder for journalists to establish the truth.
The visual of the miracle child perfectly captured this dilemma. It was designed to evoke an emotional reaction; but it also looked like the kind of footage that journalists hope to find during crisis reporting. The video was vertically filmed. The camera movement was wobbly. The lighting was grainy. The muffled background sound felt believable. The video presented none of the glaring distortions that once made AI interventions easy to spot. Fabricated videos, today, can be produced to mimic the aesthetics of authentic citizen journalism. It was for this precise reason that the floods presented journalists with a crisis that went beyond recognising AI-generated images: they were increasingly obliged to verify whether the footage reaching them was genuine in the first place.
Screenshot of video debunked by Factly
Newsmobile’s verification process increasingly involves examining individual frames, conducting reverse-image searches, analysing visual inconsistencies and using AI-detection tools where appropriate. “However, detection tools are only one part of the process. We still need corroboration,” Shukla said.
“We cannot rely solely on automated detection tools,” Sumit Dubey, South Asia digital verification editor for AFP’s Fact Check vertical, told Index. Fact-checking, he explained, requires going beyond automated outputs: tracing the original source, examining context, verifying metadata, looking for corroborating evidence and, where possible, speaking to relevant sources. “Automated tools are useful indicators, but fact-checking cannot be outsourced to a button,” Dubey said.
The proliferation of synthetic media output in the wake of natural disasters is not new. It’s the sophistication of the output that has changed. For years, much of the public conversation around AI manipulation centred on faces: celebrity deepfakes, altered political speeches and fabricated portraits. The Nepal floods flipped that narrative. Almost every fabricated clip featured rivers, glaciers, collapsing hillsides, dams, flooded valleys or stranded animals.
“AI-generated environmental content poses a different problem: [the] tells are failures of physics rather than anatomy,” Mahsa Alimardani told Index. Alimardani is associate director for technology threats and opportunities at Witness, a global organisation helping people use video and technology to protect and defend human rights
This explains why the Nepal videos were so plausible. The miracle girl video wasn’t believable just because of her facial expressions. It was convincing because the sludge moved convincingly, because the choreography resembled actual emergency rescues, because the emotional tonality seemed genuine. “Most post-hoc detection tools were designed for content featuring physical characteristics, and they currently struggle significantly with environmental and disaster content,” Bruna Martins dos Santos – policy and advocacy manager, Witness Technology Threats and Opportunities – told Index. “Journalists might also struggle because many tools have limited capabilities to identify synthetic content generated elsewhere,” she said.
Another emerging trend is the increasingly complicated role of Big Tech itself in detecting and labelling fake generative-AI content. Some of the clearest AI verdicts came from detection tools built by the very same companies whose models created the same fakes. Three weeks after the US state of California and the EU mandated the requirement for greater transparency with respect to AI-generated content, Witness and Indicator tested thirteen major AI providers. They found that seven had no public detector at all. “The capability to identify synthetic content partly existed. The governance around it did not,” Alimardani said.
When both authentic and synthetic content circulate in disaster situations, the biggest challenge that journalists face is that of establishing truth amidst widespread public uncertainty and doubt. According to Martins dos Santos, “Journalists and local fact-checkers can face immense time pressure, often wasting scarce resources re-detecting the same recycled disaster footage across multiple platforms and languages.”
The greatest danger emerging from Nepal may not be that people believed the fake videos. It may be that they have begun to doubt the real ones. Shukla’s concern is that people may stop trusting visual evidence altogether. “Imagine a situation where genuine footage of a disaster emerges, but people dismiss it as AI-generated. Facts become secondary to perception.”
Within days of the floods, even authentic CCTV footage from the Gyirong border was being perceived as AI-generated, prompting fact-checkers to establish that the scenes it depicted were real. This exposed a troubling issue: even authentic visuals can no longer be taken at face value. Martins dos Santos told Index that the phenomenon has a name: “Liar’s dividend”, whereby the mere existence of convincing synthetic content allows bad actors to dismiss genuine, damaging evidence as fake simply by claiming it was AI-generated. This means that genuine footage is no longer accorded credibility immediately; it may first be perceived with suspicion.
“This means a video may look completely authentic and still be fabricated. Conversely, genuine footage can contain compression artefacts that make it appear suspicious,” Shukla warned. Appani, for his part, noted that synthetic media doesn’t merely affect verification in real time. It also reshapes collective memory. “Widely circulated synthetic or recycled footage may eventually be remembered as genuine imagery from the disaster. Generative-AI could therefore affect not only how disasters are verified in real-time, but also how they are documented and remembered in the future.”The Nepal floods left destruction in its wake, and also a warning for journalists covering climatic disaster. The defining challenge of disaster reporting is no longer capturing the first image, but establishing which image the world can trust. In Shukla’s words: “For fact-checkers, the challenge is not just determining whether a video is real. It is establishing whether it actually belongs to the event it claims to depict.”
READ MORE](https://fsnn.net/wp-content/uploads/2026/10/indexlogodefault-1.jpg "As news broke that 84,270 people, including more than 32,000 children, had been affected by Nepal’s catastrophic floods, one visual offered something the world desperately needed in the moment: hope.
The video shows a little girl, buried under thick mud, being is rescued by Nepal’s emergency teams. Instinctively, the viewer roots for her as rescuers carry her to safety, wrapping her shivering body in a red shawl. But the most moving moment of the video comes towards the end. The girl is sitting motionless in a relief camp, an officer feeding her warm soup. As the camera lingers on her face, a faint smile almost begins to show. It is the kind of image that can offer the viewer respite from the horrors of a humanitarian crisis. The video rapidly spread rapidly across all social media platforms, with commentators calling it a miracle.
Of sorts. In fact, it was an AI miracle.
Agence France-Presse (AFP) Fact Check’s investigation led them to an India-based creator, who admitted creating the video with the help of AI-generative tools. His justification was disarmingly straightforward: emotionally engaging visuals perform well online.
This was just one of the many digitally-manufactured “miracles” put into social media heavy rotation during the floods. Another clip, showing a bridge collapsing into floodwaters swallowing vehicles and human lives, seemed to have been lifted straight out of a Roland Emmerich disaster film. The video racked up more than 2 million views on X before it was debunked as AI-generated by Factly. What made this social media post stand out for me was the identity of one of the people sharing it – an experienced news editor. The post, still pinned on his X profile and with more than 2 million views, suggests that even experienced newsroom decision-makers could be fooled by synthetic disaster footage – or, worse still, be untroubled about the veracity of such videos.
Both videos incorporate the visual language of short films, but were not released in cinemas or on Netflix. They surfaced on Facebook, X and Instagram, even as emergency teams continued to search for the thousands of people that went missing after a glacial collapse triggered devastating floods in Nepal and Tibet on August 26. Even as survivors photographed authentic eyewitness footage, synthetic visuals emerged alongside them. The floods became one of the first major climate disasters in South Asia where documentation and fabrication coexisted in the same visual ecosystem.
Screenshot of debunked Nepal floods footage
“During earlier disasters, our primary challenge was identifying old footage, establishing where a video was recorded and determining whether it was being presented in the correct context,” Saurabh Shukla told Index. Shukla is founder and editor-in-chief of NewsMobile, an India-based independent media organisation that debunks AI-related deepfakes and focuses on verified information. “The biggest difference today is the sheer speed at which fabricated and misleading visuals entered the information ecosystem,” Shukla said. “With generative AI, we now have an additional challenge – determining whether the event depicted in a video ever happened in the first place.”
This has fed into a media environment where unverified visuals are disseminated faster than the verified facts reported by journalists. “Speed without verification can amplify confusion, while verification without timely reporting can leave an information vacuum,” Shukla said. “The challenge for newsrooms is to strike the right balance between being fast and being accurate.”
The first images of rescue missions in secluded mountainous regions almost always come from locals, and thus forming the spine of disaster reporting. These inevitably blurry and shaky recordings do not have the polished quality of professional news reporting, but do carry something far more consequential: the credibility that comes with eyewitness documentation. These images often shape the core of the collective memory of a tragedy. But in an age of visual manipulation and trickery, that presumption of authenticity no longer holds.
Factly conducted 30 fact-checks into coverage of the Nepal floods; fully one-third involved AI-generated visuals. “Beyond checking the location, source, and context of a visual, we also had to establish whether the visual itself was authentic,” Akshay Kumar Appani, lead fact checker and researcher at Factly, told Index. “During breaking disasters, old AI-generated visuals can spread rapidly while being presented as current footage. This makes it harder for journalists to establish the truth.
The visual of the miracle child perfectly captured this dilemma. It was designed to evoke an emotional reaction; but it also looked like the kind of footage that journalists hope to find during crisis reporting. The video was vertically filmed. The camera movement was wobbly. The lighting was grainy. The muffled background sound felt believable. The video presented none of the glaring distortions that once made AI interventions easy to spot. Fabricated videos, today, can be produced to mimic the aesthetics of authentic citizen journalism. It was for this precise reason that the floods presented journalists with a crisis that went beyond recognising AI-generated images: they were increasingly obliged to verify whether the footage reaching them was genuine in the first place.
Screenshot of video debunked by Factly
Newsmobile’s verification process increasingly involves examining individual frames, conducting reverse-image searches, analysing visual inconsistencies and using AI-detection tools where appropriate. “However, detection tools are only one part of the process. We still need corroboration,” Shukla said.
“We cannot rely solely on automated detection tools,” Sumit Dubey, South Asia digital verification editor for AFP’s Fact Check vertical, told Index. Fact-checking, he explained, requires going beyond automated outputs: tracing the original source, examining context, verifying metadata, looking for corroborating evidence and, where possible, speaking to relevant sources. “Automated tools are useful indicators, but fact-checking cannot be outsourced to a button,” Dubey said.
The proliferation of synthetic media output in the wake of natural disasters is not new. It’s the sophistication of the output that has changed. For years, much of the public conversation around AI manipulation centred on faces: celebrity deepfakes, altered political speeches and fabricated portraits. The Nepal floods flipped that narrative. Almost every fabricated clip featured rivers, glaciers, collapsing hillsides, dams, flooded valleys or stranded animals.
“AI-generated environmental content poses a different problem: [the] tells are failures of physics rather than anatomy,” Mahsa Alimardani told Index. Alimardani is associate director for technology threats and opportunities at Witness, a global organisation helping people use video and technology to protect and defend human rights
This explains why the Nepal videos were so plausible. The miracle girl video wasn’t believable just because of her facial expressions. It was convincing because the sludge moved convincingly, because the choreography resembled actual emergency rescues, because the emotional tonality seemed genuine. “Most post-hoc detection tools were designed for content featuring physical characteristics, and they currently struggle significantly with environmental and disaster content,” Bruna Martins dos Santos – policy and advocacy manager, Witness Technology Threats and Opportunities – told Index. “Journalists might also struggle because many tools have limited capabilities to identify synthetic content generated elsewhere,” she said.
Another emerging trend is the increasingly complicated role of Big Tech itself in detecting and labelling fake generative-AI content. Some of the clearest AI verdicts came from detection tools built by the very same companies whose models created the same fakes. Three weeks after the US state of California and the EU mandated the requirement for greater transparency with respect to AI-generated content, Witness and Indicator tested thirteen major AI providers. They found that seven had no public detector at all. “The capability to identify synthetic content partly existed. The governance around it did not,” Alimardani said.
When both authentic and synthetic content circulate in disaster situations, the biggest challenge that journalists face is that of establishing truth amidst widespread public uncertainty and doubt. According to Martins dos Santos, “Journalists and local fact-checkers can face immense time pressure, often wasting scarce resources re-detecting the same recycled disaster footage across multiple platforms and languages.”
The greatest danger emerging from Nepal may not be that people believed the fake videos. It may be that they have begun to doubt the real ones. Shukla’s concern is that people may stop trusting visual evidence altogether. “Imagine a situation where genuine footage of a disaster emerges, but people dismiss it as AI-generated. Facts become secondary to perception.”
Within days of the floods, even authentic CCTV footage from the Gyirong border was being perceived as AI-generated, prompting fact-checkers to establish that the scenes it depicted were real. This exposed a troubling issue: even authentic visuals can no longer be taken at face value. Martins dos Santos told Index that the phenomenon has a name: “Liar’s dividend”, whereby the mere existence of convincing synthetic content allows bad actors to dismiss genuine, damaging evidence as fake simply by claiming it was AI-generated. This means that genuine footage is no longer accorded credibility immediately; it may first be perceived with suspicion.
“This means a video may look completely authentic and still be fabricated. Conversely, genuine footage can contain compression artefacts that make it appear suspicious,” Shukla warned. Appani, for his part, noted that synthetic media doesn’t merely affect verification in real time. It also reshapes collective memory. “Widely circulated synthetic or recycled footage may eventually be remembered as genuine imagery from the disaster. Generative-AI could therefore affect not only how disasters are verified in real-time, but also how they are documented and remembered in the future.”The Nepal floods left destruction in its wake, and also a warning for journalists covering climatic disaster. The defining challenge of disaster reporting is no longer capturing the first image, but establishing which image the world can trust. In Shukla’s words: “For fact-checkers, the challenge is not just determining whether a video is real. It is establishing whether it actually belongs to the event it claims to depict.”
READ MORE")
Listen to the article
Newsmobile’s verification process increasingly involves examining individual frames, conducting reverse-image searches, analysing visual inconsistencies and using AI-detection tools where appropriate. “However, detection tools are only one part of the process. We still need corroboration,” Shukla said.
“We cannot rely solely on automated detection tools,” Sumit Dubey, South Asia digital verification editor for AFP’s Fact Check vertical, told Index. Fact-checking, he explained, requires going beyond automated outputs: tracing the original source, examining context, verifying metadata, looking for corroborating evidence and, where possible, speaking to relevant sources. “Automated tools are useful indicators, but fact-checking cannot be outsourced to a button,” Dubey said.
The proliferation of synthetic media output in the wake of natural disasters is not new. It’s the sophistication of the output that has changed. For years, much of the public conversation around AI manipulation centred on faces: celebrity deepfakes, altered political speeches and fabricated portraits. The Nepal floods flipped that narrative. Almost every fabricated clip featured rivers, glaciers, collapsing hillsides, dams, flooded valleys or stranded animals.
“AI-generated environmental content poses a different problem: [the] tells are failures of physics rather than anatomy,” Mahsa Alimardani told Index. Alimardani is associate director for technology threats and opportunities at Witness, a global organisation helping people use video and technology to protect and defend human rights
This explains why the Nepal videos were so plausible. The miracle girl video wasn’t believable just because of her facial expressions. It was convincing because the sludge moved convincingly, because the choreography resembled actual emergency rescues, because the emotional tonality seemed genuine. “Most post-hoc detection tools were designed for content featuring physical characteristics, and they currently struggle significantly with environmental and disaster content,” Bruna Martins dos Santos – policy and advocacy manager, Witness Technology Threats and Opportunities – told Index. “Journalists might also struggle because many tools have limited capabilities to identify synthetic content generated elsewhere,” she said.
Another emerging trend is the increasingly complicated role of Big Tech itself in detecting and labelling fake generative-AI content. Some of the clearest AI verdicts came from detection tools built by the very same companies whose models created the same fakes. Three weeks after the US state of California and the EU mandated the requirement for greater transparency with respect to AI-generated content, Witness and Indicator tested thirteen major AI providers. They found that seven had no public detector at all. “The capability to identify synthetic content partly existed. The governance around it did not,” Alimardani said.
When both authentic and synthetic content circulate in disaster situations, the biggest challenge that journalists face is that of establishing truth amidst widespread public uncertainty and doubt. According to Martins dos Santos, “Journalists and local fact-checkers can face immense time pressure, often wasting scarce resources re-detecting the same recycled disaster footage across multiple platforms and languages.”
The greatest danger emerging from Nepal may not be that people believed the fake videos. It may be that they have begun to doubt the real ones. Shukla’s concern is that people may stop trusting visual evidence altogether. “Imagine a situation where genuine footage of a disaster emerges, but people dismiss it as AI-generated. Facts become secondary to perception.”
Within days of the floods, even authentic CCTV footage from the Gyirong border was being perceived as AI-generated, prompting fact-checkers to establish that the scenes it depicted were real. This exposed a troubling issue: even authentic visuals can no longer be taken at face value. Martins dos Santos told Index that the phenomenon has a name: “Liar’s dividend”, whereby the mere existence of convincing synthetic content allows bad actors to dismiss genuine, damaging evidence as fake simply by claiming it was AI-generated. This means that genuine footage is no longer accorded credibility immediately; it may first be perceived with suspicion.
“This means a video may look completely authentic and still be fabricated. Conversely, genuine footage can contain compression artefacts that make it appear suspicious,” Shukla warned. Appani, for his part, noted that synthetic media doesn’t merely affect verification in real time. It also reshapes collective memory. “Widely circulated synthetic or recycled footage may eventually be remembered as genuine imagery from the disaster. Generative-AI could therefore affect not only how disasters are verified in real-time, but also how they are documented and remembered in the future.”
The Nepal floods left destruction in its wake, and also a warning for journalists covering climatic disaster. The defining challenge of disaster reporting is no longer capturing the first image, but establishing which image the world can trust. In Shukla’s words: “For fact-checkers, the challenge is not just determining whether a video is real. It is establishing whether it actually belongs to the event it claims to depict.”
Read the full article here
Fact Checker
Verify the accuracy of this article using AI-powered analysis and real-time sources.