The footage going round after a disaster shows the disaster: clips circulating in the first hours after a flood, quake or landslide document the event they are captioned with
By The Conspiratory EditorsAugust 31, 2026
Where the evidence lands: ContradictedThat video circulating widely in the hours and days after a major natural disaster, particularly video that shows no signs of AI generation or digital editing, can be treated as documentary footage of that disaster and shared on that basis.
Contradicted. This rates a reflex rather than any one clip, and the Nepal flood of 26 August 2026 is the cleanest test of it yet recorded. In the six days after the water came down, six fact-checking organisations in four countries published roughly twenty separate debunks of material presented as footage of that flood, and one authentication. The single item rated authentic, CCTV of the Gyirong Port border crossing being destroyed, was the one an audience of millions had decided was AI-generated. So the reflex fails in both directions at once: fabricated and recycled clips are accepted because they carry no visible tells, and real evidence is rejected because it looks the way catastrophe actually looks. Nothing here disputes the disaster, which is documented and severe. What is rated is the inference that a clip circulating during a catastrophe is footage of it.
Believed by: Almost everyone during the first hours of a disaster, including careful people and working journalists. Urgency, sympathy and the wish to warn others all push toward sharing before checking, and the checks most people know how to run do not test the thing that is false
A second Indonesian television drama has been mistaken for the Nepal floods, one day after the first, and the pairing says something the single case did not. On 3 September Fact Crescendo rated false a video circulating as devastation from the flooding, showing debris, bodies and a woman with a small boy. Reverse-searching a keyframe found it posted on 3 July by an Indonesian account promoting Lautan Cinta, a serial broadcast daily on SCTV, and a second clip from the same production surfaced on a Facebook page on 28 July. The detail worth keeping is where the tell was. Fact Crescendo says its first suspicion came from the footage itself: a filming crew is visible in one part of the shot. The clip travelled anyway. So did the context, in the other direction. The original Indonesian caption was a thank-you to the crew for hours spent in cold water to get the scene right, which means the video was published as explicitly behind-the-scenes material and only became disaster footage once that caption was stripped. Nothing was faked to make this work. Two serials, two fact-checkers and two days suggest scripted flood scenes from Indonesian television are becoming a standing supply for this genre, which is a supply-side pattern rather than a run of coincidences, and it is worth watching after the next flood anywhere. source →
A check Full Fact published this afternoon supplies the mirror image of the provenance problem this file has been describing. A clip of a wave demolishing a street, shared thousands of times as footage of the flood, is not merely miscaptioned: it is wholly generated, and Full Fact identified it by an invisible SynthID watermark showing it came from Google's AI. Set that against the pattern in the entries above, where nearly every item was authentic footage carrying a false caption, and where no signature scheme could have helped because there was nothing wrong with the file to detect. Here the reverse holds. A generated clip carried its generator's mark, and the mark survived the journey. Provenance tooling works where the fabrication is in the pixels and fails where the fabrication is in the caption, which is most of this file. Full Fact lists the ordinary tells alongside the watermark, red containers and motorbikes appearing and vanishing and people walking towards the wave rather than away from it, and puts the flood's toll at over 1,000 killed, which independently confirms the figure corrected in the entry above. source →
A week on, the disaster is roughly twice the size of the figures this file was built around. Al Jazeera reported on 2 September, citing Reuters, that 1,114 people have been killed in Nepal with nearly 4,000 missing, alongside 16 dead and 546 missing in the Tibet Autonomous Region, against the 547 to 600 that fact-checkers recorded in the first two days and that are quoted above. The UN Development Programme put the debris at 2.2 million tonnes on a preliminary estimate; eleven hydropower stations in the affected area have stopped operating, more than 270 workers have been pulled from tunnels beneath them, and around 600 are still unaccounted for. The early figures are left standing where they appear above, because what was known in the first 48 hours is part of the argument here: the fake footage attached itself to numbers that understated the event by half. That is worth saying plainly rather than quietly correcting. A page that rates other people's handling of this disaster should not itself leave a reader with half the death toll. source →
The elephant clip has now split the fact-checkers, which is a first for this set. Full Fact, cited further down this page, treated the video of an elephant carrying a man out of floodwater as genuine footage that simply was not from Nepal. Snopes reached a different conclusion on 1 September: unable to find an original, it dated the earliest copy to a YouTube upload of 5 August, three weeks before the Nepal floods began on 26 August, and judged the footage itself AI-generated on visual grounds, listing harsh shadows, inconsistent skin texture, object bleeding and misshapen extremities. It also found that one widely shared version carries Hindi overlay text and Thai speech, which places it in neither country the caption claims. Snopes is explicit that it could not definitively confirm authenticity either way, so this is a disagreement rather than a correction. It matters for the argument this file makes. The mechanisms catalogued here were treated as separable, generated from nothing on one side and real footage of the wrong event on the other, and here two careful organisations examining the same clip cannot agree which one applies. A further layer sits underneath it: commenters point to a real 2016 video from the Elephant Nature Park in Chiang Mai, in which an elephant named Kham Lha went into a river towards her caretaker, and Snopes confirms that story is real while the viral clip is not that video. A true story is orbiting the false one, close enough to make the false one feel verified. source →
Day seven, and three more landed this morning, which settles the question of whether this was a first-week surge that would burn itself out. Fact Crescendo traced a clip billed as the first footage of the Nepal floods to Humantay Lake in Peru, adding a fifth country to the borrowed-source list alongside the Congo, Japan, Indonesia and India. Factly rated two: old flood visuals from Kei Panyor district in Arunachal Pradesh circulating as Nepal's Rasuwa area, and a video of people pulling 500-rupee notes out of floodwater, which turns out to be scripted and shot in Uttar Pradesh. Two things are worth adding to the record. The first-footage framing is a claim of primacy rather than of content, and it is a strong hook precisely because it promises something nobody else has, so it invites sharing before checking in a way that an ordinary caption does not. And the money video belongs with the staged women in the entry above: not recycled, not generated, but performed for the camera and then relabelled. That category now has two instances a week apart, which makes it a habit rather than an outlier. Note also what the Arunachal Pradesh item does to the geography argument. Kei Panyor is in India, the flood is in Nepal, and both are Himalayan; a viewer applying the sensible test of whether the terrain looks right would pass it. source →
The pattern repeated on a second disaster within days, and the repeat sharpens the finding. Fact Crescendo rated a video of a flooded river in hill country that was circulating as footage of a flood in Himachal Pradesh. It is the Grand Canyon. A flash flood tore through a Grand Canyon trail on Saturday 29 August 2026, sending a bunkhouse drifting downstream; the BBC reported dozens of hikers airlifted to safety with about 15 unaccounted for, and FOX Weather posted the clip to its own Instagram account on 30 August with an accurate caption. Within roughly a day it was in India as Himachal Pradesh. Two things follow. First, the recycling window has collapsed. Most of the Nepal material was pulled from 2023 and 2024, which suggested an archive being raided; this was one day old, which means the pool is not the past but the present. Second, and more useful, two simultaneous disasters on different continents now contaminate each other's imagery, so the recycled footage is itself real footage of a real catastrophe with real casualties, just the wrong one. Note where the clip came from: a working news organisation, on its own account, correctly captioned, publicly checkable. The provenance chain was clean and someone stripped it anyway, which is the argument in the closing section of this file arriving faster than expected. source →
Fact Crescendo was still filing on this event five days after the flood, rating a child rescue video from 2024 that had begun circulating as Nepal footage. Worth noting for what it says about duration: the recycled-clip problem does not resolve when the news cycle moves on, because the supply of authentic disaster footage from previous years is effectively unlimited and each new wave draws on it again. source →
The full story
Twenty debunks and one authentication
The flood came down the valley at the Nepal-China border on 26 August 2026 and destroyed the Gyirong Port crossing. Two days later Full Fact recorded more than 547 dead. CBS News, cited by Snopes, put it at almost 600 killed with roughly 2,500 missing. Those are the facts of the event, and nothing in this file disputes them.
What happened next is the subject. Over six days, six fact-checking organisations across four countries published roughly twenty separate debunks of material presented as footage of that flood: Lead Stories seven, Fact Crescendo nine, Full Fact two, Factly two, Boatos one. In the same window exactly one item was rated authentic, by Snopes.
The ratio is not the interesting part. The identity of the authenticated item is. It was the CCTV of the water hitting the border gate, the single most widely seen piece of evidence from the disaster, and it had already been convicted in public of being generated.
The one they were sure was fake
The clip shows floodwater cascading down a valley while dozens of people run, then crashing onto the border port and burying the frame in mud. It spread across X, Facebook and Reddit. Under one X post carrying it, which passed thirteen million views, commenters said it was AI-generated, and enough readers absorbed that to start arriving at Snopes searching for confirmation that the video was fake.
It was not fake. Snopes found no evidence of generation or doctoring, the Associated Press and other reputable outlets had independently deemed the footage authentic, and drone footage taken afterwards confirmed the structure had been pulverised. Snopes was still trying to locate the original CCTV source as it published, and said so, which is the correct way to report an unfinished thread.
Consider what the accusation required. Not evidence, and not a tell: the clip was rejected because a wall of water demolishing a gate looks like something a model would produce. That instinct will misfire hardest on the most extreme real events, because those are the ones that look least like ordinary footage. An audience trained to distrust the spectacular will be most confident exactly when the record matters most.
It is not one failure, it is five running at once
Treating disaster misinformation as a single problem called fake videos obscures how differently the pieces work. In this one event, at least five distinct mechanisms ran side by side.
Generated from nothing. A bridge collapsing into floodwaters, rated AI-generated by Lead Stories. Before-and-after damage images, made with OpenAI tools. This is the category everyone is watching for, and it was a minority of the total.
Real footage, wrong disaster. A rescue from the Congo in 2023. A landslide in Japan. An excavator swept away in Chamoli, Uttarakhand. A man chest-deep in debris in Prayagraj. A Buddha statue and a child rescue, both from 2024. All authentic video of real events, which is why they survived every check a careful person knows how to run.
Fiction, relabelled as record. The scene of a woman and child in a flooded area came from an Indonesian television series. It was never a claim about the world at all until somebody attached one.
Performance, presented as documentation. A staged video of women acting dead, pushed as flood victims in Nepal and, in parallel, in Assam. The same artifact serving two different disasters tells you the material was chosen for effect rather than for any connection to a place.
Causal claims wearing footage as evidence. HAARP caused it. An F-16 caused it. These are assertions about mechanism, and they are not answered by authenticating or debunking a clip. The reply is physical: this was a glacial and landslide event in the Himalayas, and a transmitter that aims high-frequency energy upward at the ionosphere is not a candidate cause of a mountain flood.
Six organisations, six fragments
Every item above was caught. That is the part worth sitting with, because the failure was not one of detection. Lead Stories rated seven, Fact Crescendo nine, Full Fact two, Factly two, Boatos one, Snopes one. The overlap between them amounts to a single elephant video that Lead Stories and Full Fact both happened to examine.
So no organisation saw more than about half the set, and no reader following any one of them saw the shape of it. A person reading Full Fact learned that HAARP did not cause the flood and that an elephant video was old. A person reading Fact Crescendo learned that nine clips came from India, Japan, Indonesia and 2024. Neither learned that roughly twenty pieces of material were circulating, that the sources spanned four countries and a television drama, or that the one authentic clip was the one being denounced.
The corrections existed and the aggregate did not, which is a different problem with a different fix, and it is not more fact-checking.
What actually works in the first six hours
The lesson of this week is not scepticism, because scepticism was abundant and it produced the worst single error in the set. Thirteen million people were shown the best evidence of a disaster and told, persuasively, to disbelieve it.
What separated the true item from the false ones was never appearance. It was provenance: a chain running back to a named source through parties who can be checked. The Gyirong Port CCTV had the Associated Press behind it and drone footage of the same wrecked structure taken afterwards. The Congo clip, the Japanese landslide and the Indonesian drama all had provenance too, and theirs led somewhere other than Nepal. The question that sorts them is not “does this look generated” but “where did this come from, and who says so”.
That test is slower, and it is the only one that worked here. It also has an uncomfortable implication for the tooling being built to solve this. Cryptographic provenance labelling, which attests that a file is unedited and records where it originated, would have certified almost every clip in this set, because they are authentic recordings. It would have said, correctly, that the Congo rescue is a real unedited video. Nobody was ever confused about that. They were confused about which country they were looking at, and that is a claim made in a caption, where no signature reaches.
What's still unexplained
- Whether provenance labelling of the kind platforms are building would have helped here at all. Most of this material carries clean provenance, because it is authentic footage of other events; a cryptographic signature attesting that a video is unedited would have certified the Congo rescue as genuine, which it is.
- Whether the reflex to shout AI now produces more error than it prevents during disasters specifically. On this event it produced one very large error against the single most important piece of evidence, while the actual fabrications spread on other grounds entirely.
- How far the corrections travelled against the originals. One X post carrying the authentic CCTV passed thirteen million views; no debunk in this set is likely to have come close, and none of the outlets publishes reach figures that would let anyone check.
- Whether pooling would have worked. Six organisations each saw a fragment and none saw the whole, in a case where the shape of the thing only becomes visible in aggregate. Whether a shared live register during a major disaster is practical, or merely obvious in hindsight, is untested.
Point by point
The claim: The clip shows no AI artefacts and survives a reverse image search into real footage, so it documents this disaster.
What the record shows: Most of the material debunked here was real footage of real suffering, which is exactly why it passed. The Congo rescue happened. The Japanese landslide happened. The Chamoli excavator was genuinely swept away. A reverse image search that terminates at authentic footage has answered the question of fabrication and has not touched the question of when and where. The one item in this set that a detection tool would have flagged, the AI bridge collapse, is a minority of the total.
The claim: Footage that looks implausibly cinematic during a disaster is probably generated.
What the record shows: This is the failure the Gyirong Port CCTV records. A wall of water demolishing a border gate while people run looks like a rendering, and millions concluded it was one. Snopes, the Associated Press and post-event drone footage all say otherwise. Catastrophes look catastrophic, and an instinct calibrated to implausibility will reject the real record of the worst events most confidently of all.
The claim: If something this widely shared were false, a fact-checker would have flagged it by now.
What the record shows: They did, and the flagging was fragmented. Lead Stories published seven items, Fact Crescendo nine, Full Fact two, Factly two, Boatos one and Snopes one. The overlap between them is small, essentially a single elephant video that Lead Stories and Full Fact both examined. No single organisation saw more than about half the set, and no reader following any one outlet would have seen the pattern. The debunks existed; the aggregate did not.
The claim: The fabricators are anonymous accounts trying to deceive people about the disaster.
What the record shows: Often the origin is not deceptive at all and the deception is added downstream. The F-16 clip had been online since May 2026 carrying an AI watermark, which is a disclosure. The Indonesian material is a television drama, which never claimed to be news. The 2024 Buddha statue video was a real video of a real thing in 2024. In each case someone stripped a context that was publicly available and attached a new one, which is the same mechanism documented for miscaptioned video generally.
The claim: The argument is about which videos are real, so verifying footage settles it.
What the record shows: Several of the loudest claims were causal rather than documentary, and no amount of footage verification touches them. HAARP was said to have caused the flood; an aircraft was said to have caused it. Those assert a mechanism, and the answer to them is physical rather than forensic: the flood was a glacial and landslide event in the Himalayas, and a high-frequency transmitter aimed upward at the ionosphere is not a candidate cause. Treating a mechanism claim as a footage question concedes the frame.
Timeline
- 2026-08-26A landslide and flash flood come down the valley at the Nepal-China border, destroying the Gyirong Port crossing. Casualty figures move quickly over the following days: Full Fact records more than 547 dead as of 28 August, while CBS News, cited by Snopes, reports almost 600 killed and roughly 2,500 missing.
- 2026-08-27The first wave of checks lands, and every one concerns footage from somewhere else. Fact Crescendo traces a massive landslide video to Japan, a flood scene to Dharali in India, and a man trapped chest-deep in debris to Prayagraj. Lead Stories and Full Fact independently trace a video of an elephant rescuing a man to a post published weeks earlier carrying an India locator.
- 2026-08-27Lead Stories examines a clip presented as an F-16 causing the floods. The first segment has been online since May 2026 and carries an AI watermark. The claim is not only that the footage is of Nepal but that an aircraft caused a glacial flood, which is a causal assertion the footage could not establish even if it were genuine.
- 2026-08-28The heaviest day. Lead Stories rates an AI-generated bridge collapsing into floodwaters, before-and-after damage images made with OpenAI tools, rescue footage actually shot in the Congo in 2023, and a staged video of women acting dead that is circulating as flood victims in both Nepal and Assam.
- 2026-08-28Fact Crescendo publishes four more: a Buddha statue said to have survived the flood, which is a 2024 video; an excavator swept away, filmed in Chamoli, Uttarakhand; a rescue-work clip unconnected to Nepal; and a scene of a woman and child in a flooded area taken from an Indonesian television series. That last one is not recycled news footage at all. It is fiction, repurposed as documentary evidence of a real death toll.
- 2026-08-28Full Fact rejects the claim that HAARP caused the flood, noting that the facility is a high-frequency transmitter for studying the ionosphere and thermosphere, the atmospheric layers beginning around 37 miles up, and cannot control weather or natural disasters. Factly publishes a compilation piece establishing that none of a set of widely shared visuals relates to the floods.
- 2026-08-28Snopes rates a video True. The CCTV showing floodwaters cascading down the valley and obliterating the Gyirong Port gate is authentic; the Associated Press and other outlets had also deemed it genuine, and drone footage of the aftermath confirmed the structure was pulverised. Commenters on one X post carrying it, with more than thirteen million views, had insisted it was AI-generated, and Snopes notes readers were arriving at the site searching for the clip precisely because they had been told it was fake.
- 2026-08-29Boatos, in Brazil, rates a story that a house left standing after the mudflow belonged to a pastor. The providence narrative is a distinct genre from the recycled clip, and it attaches to disasters reliably enough to be worth naming separately.
- 2026-08-31Fact Crescendo is still filing on the same event five days later, rating a child rescue video from 2024 that is circulating as Nepal footage. By this point the tally across six organisations in four countries stands at roughly twenty debunks and one authentication.
Contradicted. This rates a reflex rather than any one clip, and the Nepal flood of 26 August 2026 is the cleanest test of it yet recorded. In the six days after the water came down, six fact-checking organisations in four countries published roughly twenty separate debunks of material presented as footage of that flood, and one authentication. The single item rated authentic, CCTV of the Gyirong Port border crossing being destroyed, was the one an audience of millions had decided was AI-generated. So the reflex fails in both directions at once: fabricated and recycled clips are accepted because they carry no visible tells, and real evidence is rejected because it looks the way catastrophe actually looks. Nothing here disputes the disaster, which is documented and severe. What is rated is the inference that a clip circulating during a catastrophe is footage of it.
Reviewed by The Conspiratory Editors · Last reviewed September 3, 2026 · How we rate
Common questions
Is Disaster footage true?
Contradicted. This rates a reflex rather than any one clip, and the Nepal flood of 26 August 2026 is the cleanest test of it yet recorded. In the six days after the water came down, six fact-checking organisations in four countries published roughly twenty separate debunks of material presented as footage of that flood, and one authentication. The single item rated authentic, CCTV of the Gyirong Port border crossing being destroyed, was the one an audience of millions had decided was AI-generated. So the reflex fails in both directions at once: fabricated and recycled clips are accepted because they carry no visible tells, and real evidence is rejected because it looks the way catastrophe actually looks. Nothing here disputes the disaster, which is documented and severe. What is rated is the inference that a clip circulating during a catastrophe is footage of it.
What is Disaster footage?
On 26 August 2026 a landslide and flash flood struck the Nepal-China border. Full Fact recorded more than 547 dead two days later; CBS News, cited by Snopes, put the toll at almost 600 with around 2,500 missing; a week later Reuters put the Nepali dead at 1,114, with nearly 4,000 missing across both sides of the border. Within hours the…
What does the evidence show?
Most of the material debunked here was real footage of real suffering, which is exactly why it passed. The Congo rescue happened. The Japanese landslide happened. The Chamoli excavator was genuinely swept away. A reverse image search that terminates at authentic footage has answered the question of fabrication and has not touched the que…
Why do people believe it?
Demand for images vastly outruns supply in the first hours. A disaster in a remote border valley generates enormous appetite for pictures and almost no pictures, and recycled material rushes into the vacuum because something has to fill it.
What is still unresolved?
Whether provenance labelling of the kind platforms are building would have helped here at all. Most of this material carries clean provenance, because it is authentic footage of other events; a cryptographic signature attesting that a video is unedited would have certified the Congo rescue as genuine, which it is.
Sources
- 1.Is This Video of a Flood Destroying a Nepal-China Border Crossing Real?, Snopes (2026)
- 2.HAARP didn't cause the flood on the Nepal-Tibet border, Full Fact (2026)
- 3.Clip of elephant rescuing man does not show this week's flood in Nepal, Full Fact (2026)
- 4.Fact Check: FAKE Video Shows Nepal Bridge Collapsing Into Floodwaters, It Is AI-Generated, Lead Stories (2026)
- 5.Fact Check: FAKE Before And After Images Of Nepal Flood Damage, Made With OpenAI Tools, Lead Stories (2026)
- 6.Fact Check: FAKE Video Does NOT Show F-16 Causing Nepal Floods, First Clip Has Been Online Since May 2026 With AI Watermark, Lead Stories (2026)
- 7.Fact Check: Video Does NOT Show Nepal Flood Rescue Of People Trapped In Mud, It Is From The Congo In 2023, Lead Stories (2026)
- 8.Fact Check: Staged Video Shows Women Acting Dead, Misrepresented As Flood Victims In Nepal And Assam, India, Lead Stories (2026)
- 9.The Viral Video Of A Woman And Kid In A Flood-Hit Area Is Unrelated To Nepal Floods, Fact Crescendo (2026)
- 10.Old video of massive landslide in Japan goes viral linking to recent Nepal flood, Fact Crescendo (2026)
- 11.None of these visuals are related to the Nepal floods in August 2026, Factly (2026)
Embed this verdict on your site
Paste this snippet to show our sourced verdict as a small card, with a link back to the full case file. Free to use.
<iframe src="https://theconspiratory.com/embed/disaster-video-misinformation" title="The Conspiratory verdict" width="520" height="190" style="border:0;max-width:100%" loading="lazy"></iframe>Help us investigate
This is a living case file. If you spot an error or know evidence we missed, tell us, and weigh in on where you land.
Where do you land?
Cast your read on this one.
Comments
Add your take. Comments are read and approved by a human before they appear, so keep it on topic and civil. Please do not accuse named, living people of crimes.