The video going round proves what happened: authentic footage circulating during a breaking news event documents the event it is captioned with

Snopes rated a case on 31 August that runs this file's mechanism on a quotation rather than on footage, and the rating it earned is the interesting part: Correct Attribution. Posts circulating in August 2026 said a Florida meteorologist had gone off-script during a broadcast to say he feared he could no longer predict hurricanes, because of 'the cuts, the gutting, the sledgehammer attack on science in general.' John Morales of NBC 6 South Florida did say exactly that, on camera, in the words quoted. He said it on 3 June 2025, in a clip posted to his station's own YouTube channel, after re-airing 2019 footage of himself discussing Hurricane Dorian. The August posts carried the implication that he was speaking about the season now underway. Snopes has asked him whether he would still say it and had no reply at the time of writing. Nothing was edited, nothing was generated, and nobody was misidentified. The only false element is the year, and it is doing all the work, because a forecaster's warning about the coming hurricane season means something different from the same warning made fifteen months earlier. Readers checking this the usual way would have confirmed it at every step: the speaker is real, the station is real, the video is real, the words are exact. Note what that costs. The corpus already documents the same displacement applied to photographs, including the picture of Trump in Moscow dated to 1987 when it was taken in 1996. A quotation carries even more apparent verification with it than an image does, since it can be traced to a named person on a named broadcast, and none of that provenance touches the question of when. source →

This file argues that the falsehood lives in the caption, and here is a case that extends the argument by moving it somewhere else: the cut. Africa Check reported on 25 August that a video circulating on Facebook, presented as Uganda's chief of defence forces Muhoozi Kainerugaba repeatedly struggling to get through a public address, has been edited. A reverse image search led to the original, footage of him speaking at an event connected to the African Union Support and Stabilisation Mission in Somalia. Comparing the two shows jump cuts removing parts of his remarks. In the unedited recording he speaks clearly, with occasional pauses and fillers; stripping the material between them makes those pauses dense and the delivery disjointed. Nothing was added. No generated voice, no fabricated face, no counterfeit branding, and the caption does not misplace the footage in time or space, because the event and the speaker are exactly what they appear to be. Every frame is authentic and in its original order. That leaves nothing for a detection tool to find and no false statement to check, only a false impression assembled out of true material by deletion. The payload arrives in the accompanying text, which draws a conclusion about the speaker's competence that the edit exists to evidence. Worth holding onto as a category: subtractive manipulation is cheaper than any technique logged elsewhere on this site, it is invisible to every check aimed at what has been inserted, and the only way to catch it is to find the original and watch both. source →

A still photograph rather than a clip, and a shift in date rather than in place, which together open an axis this file has not covered. Snopes rated Miscaptioned on 22 August a black-and-white image circulating since at least 18 August as the US president in Moscow in July 1987, aged 41, at the invitation of Yuri Dubinin, then the Soviet ambassador to the United States. The photograph is authentic and it does show him in Moscow. It was taken in November 1996, when he was 50 and the Soviet Union no longer existed, reportedly on a second trip surveying real estate sites. Dubinin was not ambassador by then. Here is why the misdating is not incidental colour. The 1987 visit is real and documented: his own 1987 book records the Dubinin invitation and the trip. So an authentic photograph of a real second visit is being offered as an illustration of a real first visit, and the nine-year slip is the only thing attaching the image to the Soviet-era framing that the surrounding narrative rests on. Every earlier entry here describes a caption moving footage to a different place. This one shows a caption moving an image to a different moment in the same person's life, where both moments are documented and only the pairing is false. Nothing in the photograph can register the difference, and no authenticity check has anything to report, because the picture is exactly what it appears to be. This file rates the caption, not the wider claims that travel with it. source →

A fourth continent, and a variant of the technique the entries below do not cover. Africa Check reported on 21 August that a clip circulating since July, presented as Nigerians marching over the deportation case of the beauty contestant Chidimma Adetshina in South Africa, shows nothing of the kind. The footage is of real demonstrations in Nigeria set off by a death reported under mysterious circumstances, an entirely separate matter, and no reputable outlet has reported any protest in Nigeria linked to Adetshina in August 2026. Rated false. What distinguishes this from the earthquake clips above is what gets swapped. Those were real footage of one disaster relabelled as another disaster, a substitution of place and date. Here the crowd, the country and the anger are all authentic and what has been replaced is the cause: a genuine protest is detached from what people were actually protesting about and reattached to an unrelated legal dispute involving a named individual. Nothing needs editing, because a crowd does not carry its own subject. Worth noting too that the demand ran ahead of the check. Adetshina was already among the most searched names in the United States on 20 August, a day before any fact-checker addressed what was circulating about her. source →

Later the same day, Full Fact published a second check on the Indonesian earthquake that complicates the entry below in a way worth recording. This one examines a six-clip compilation circulating on Facebook and Instagram as damage from the 15 August quake. Three of the six are genuine Indonesian footage. The other three are not: two are the October 2025 Philippines earthquake again, including the room of young people sheltering under chairs, and the first clip, showing a brown building leaning on its side, is the 7.4 magnitude Taiwan earthquake of April 2024. So the artifact is half true, and that is the point. Every piece of verification advice in circulation, including the advice this file gives, assumes the unit under test is a clip. Against a compilation that assumption breaks: a reader who spot-checks one video has a fifty per cent chance of landing on real Indonesian footage, confirming it, and treating the confirmation as covering the whole montage. The genuine clips are not padding, they are the credential. This is the first instance recorded here where authentic material is doing the work of laundering recycled material inside a single file, and it costs the person assembling it nothing beyond choosing what to bundle. source →

A textbook instance of this file's core mechanism, and it comes with a receipt for how long a single clip can stay in service. A 7.7 magnitude earthquake struck Indonesia on 15 August 2026, killing at least 70 people. Within days, footage of a room of young people swaying and sheltering under chairs, and a separate clip of a corridor ceiling giving way, spread with captions naming the Indonesian quake and its tsunami warning. Full Fact traced both on 19 August to Davao City in Mindanao during the 7.4 magnitude earthquake that hit the Philippines on 10 October 2025. Nothing about either clip was altered. What makes this one worth logging beyond the count is the chain: the same footage was already recirculated in June 2026 against a different Philippines earthquake, so it has now been attached to three separate disasters in ten months, twice in the wrong country. A clip that reads as generic shaking has no features tying it to a place, and that is precisely what gives it an indefinite shelf life. Each new quake supplies a fresh caption and the footage needs no editing at all. source →

The technique this file describes has been upgraded, and the upgrade is worth recording carefully. Full Fact reported on 18 August on a clip shared thousands of times across Facebook, Instagram and X in which a migrant appears to say, in English, that he wants money, a free house and an English woman. The source footage is real and is not obscure: the Associated Press filmed it in February 2024, and it shows a Sudanese refugee in Tunisia responding to the Albanian parliament ratifying a migration deal with Italy. He is speaking Arabic, AP reports him saying that they need safety and security because they do not know what will happen to them in Albania, and he does not mention England at all. Full Fact matched the start of the edited clip to timecode 01:00 of the original. Here is what makes it different from every other item above. In the cases collected so far, the footage was authentic and only the caption lied, which is why the advice to check the image rather than the label worked. In this one the audio is fabricated and so is the picture: the man's mouth and gestures were altered to fit the invented speech, and even the positions of the other men around him were changed. The recycled clip supplies the provenance, the generated layer supplies the words, and checking whether the video is real returns the wrong answer, because it is real and it has been made to say something its subject never said. The accompanying captions pushed a mass-deportation line. The fabrication is the story here; the words put in his mouth are not repeated. source →

The technique has moved off video, and the cleanest specimen yet is not footage at all. Lead Stories checked a post of 9 August sharing an image of a BBC news article reporting that the US Department of Agriculture had listed Wakanda, the fictional nation of the Black Panther films, as a free-trade partner, captioned to the effect that Americans should welcome their new trading partner. Every element of the artifact is real. The BBC published that article, under the headline that the US government had listed a fictional nation as a trade partner, and the USDA listing genuinely existed. Two things are missing rather than false. The article was published on 19 December 2019, and the image shared to X does not show a publication date; and the listing was a testing error rather than any belief on the agency's part that Wakanda exists. Lead Stories rated it not false but 2019 Article, which is the right verdict and an unusual one. Nothing here was fabricated, nothing was edited, and there is no synthetic media to detect. What was removed was the date, and the date was carrying the whole meaning. That is the operation this file describes, applied one step upstream: not authentic footage of an event recaptioned, but the authentic news report about the event, screenshotted with its timestamp cropped away. It is worth noting the same trap works on people checking in good faith. A reader who verifies that the BBC really did publish this finds it confirmed and stops, exactly as a reader who confirms a video is not AI-generated stops. source →

Two items on 6 August extend the labelling point above in a direction it did not cover, and add a category this file has been circling without naming. Snopes rated fake a video of an Amazon delivery robot trundling off a pier into the sea, shared on Instagram with the explanation that an isolated navigation error was to blame, and onward to Facebook, Threads, TikTok and X. Its evidence included a label on the original post declaring the clip AI-generated. Separately, Lead Stories checked a video posted to X on 5 August as a passenger plane making a fiery emergency landing on a rain-slicked Georgia highway; Hive Moderation rated it 96.1 percent likely AI, and searches turned up no such incident, only real and unrelated small-aircraft landings. Take the robot first. The entry below argues that labels do not survive redistribution, using a fabricated political video where someone had a motive to remove one. The robot clip carries no smear, accuses nobody and is essentially a joke, and its label came off just the same. That widens the finding: redistribution strips provenance regardless of anyone's intent, which is a harder problem than bad actors. The plane video names the category. These are not real events recaptioned and not attacks on anyone. They are invented incidents, minor enough to be plausible and dull enough to go unchecked, filling the space where breaking news is supposed to be. source →

A variant this file had not yet recorded surfaced on 5 August: not misdated footage and not AI, but a scripted comedy sketch relabelled as documentary evidence. Lead Stories traced a clip circulating on X from 3 August, captioned as showing a transgender woman arriving at a neighbour's door with guns after being misgendered, to the social-media creator Spicy Jayy, who published it on YouTube on 29 November 2025 as a performed skit with a small recurring cast. The identification is physical rather than forensic: the apartment door and its Ring camera appear in several of the creator's other videos, with a matching green doormat and apartment number plaque across the hall. No such confrontation happened. Two things make it worth logging here. The caption is a fabricated smear aimed at a group rather than an error about a news event, which is a harder thing to correct once it has travelled. And the version on X carried an added Russia Today watermark and came from an account styled as an RT visuals feed, so authored fiction reached an audience dressed as wire footage with the fiction stripped off. source →

A legal backdrop appeared behind this whole subject on 2 August 2026, when the EU AI Act's transparency obligations took effect under the commencement schedule in Article 113. Article 50 requires anyone deploying an AI system that generates or manipulates image, audio or video constituting a deep fake to disclose that the content is artificially generated, with an exemption for law-enforcement use and reduced disclosure where the material forms part of an evidently artistic, creative, satirical or fictional work. It is a duty to label, not a power to remove. Its central weakness is on display in a case checked two days earlier: a fabricated video of the New York mayor carried an AI disclaimer in one version and was reposted on 1 August without it. Labels attach at the point of generation or upload, and nothing yet carries them across a download and a repost somewhere else, which is the exact route most of the material in this file travels. source →

A third technique from the same news event, and this one launders the fake through a trusted brand. On 3 August an Instagram post shared an image of a Moroccan police officer apparently about to bring a rock down on a migrant climbing a rocky embankment near Ceuta, over a long humanitarian caption and the credit line 'Photo: Reuters'. Reuters published no such photograph. Lead Stories reports that OpenAI's Verify tool detected a SynthID watermark originating from OpenAI tools in the image, and notes the picture closely resembles a scene that really did occur in or near Ceuta and was recorded from several angles. So the sequence here is a synthetic re-creation of a real, documented moment, credited to a wire service that never carried it. Two things follow for this file. First, an AI check would have caught this one, which is a reminder that the argument here is against relying on a single test, not against detection. What sold it was the Reuters credit and the resemblance to something that genuinely happened. Second, note the politics: the antisemitic caricature recorded below came from an anti-migrant account and this image carries a humanitarian framing, yet both were made with the same company's tools and both were caught by the same watermark. The technique has no politics. source →

A case published hours after this file went up shows the two techniques combining, which the file had treated as alternatives. On 1 August an X post paired authentic grainy video of a woman who appeared to wave at people crossing into Ceuta with a close-up image presented as her face. Lead Stories found the close-up was AI-generated, with ZeroGPT at 91 percent, Sightengine at 99 percent, Hive Moderation at 99.8 percent, and OpenAI's verification tool reporting it was made with OpenAI tools; the earliest copy traced to a 4chan post whose caption used the word prompt. The real footage was doing the work the caption usually does, lending an authentic setting to a fabricated element, and a viewer checking the video would have found nothing wrong with it. Recaption and fabrication are not a clean either-or, and a hybrid defeats a check aimed at only one of them. source →