AI-generated videos are no longer easy to dismiss as crude clips with distorted faces, strange hands, or robotic voices. Some of the most widely shared fakes now look convincing enough to pass a casual viewing, particularly when they are attached to a story that people already expect to be true.
In recent months, fabricated videos have falsely shown politicians making inflammatory remarks, celebrities appearing to endorse products, public figures promoting financial schemes, and supposed news footage depicting events that never happened. Some have accumulated millions of views before fact-checkers or officials intervened.
The problem is becoming harder because modern AI can reproduce a familiar face and voice while leaving fewer obvious visual clues. That makes the real test less about spotting one strange frame and more about establishing where a video originated, whether credible sources confirm its claims, and whether the footage survives closer scrutiny.
The latest examples show just how quickly a convincing fake can move from social media to public discussion.
Politicians and Public Figures Are Becoming Prime Targets
Several recent cases show how political deepfakes can exploit a familiar face and a real news event.
In India, the Press Information Bureau's Fact Check Unit recently debunked AI-generated videos falsely attributed to Defence Minister Rajnath Singh, Commerce and Industry Minister Piyush Goyal, and Education Minister Dharmendra Pradhan.
The clips purported to show ministers making threatening remarks about student protests, but the government said the statements were fabricated.
A separate fake video involving Lieutenant General Rajeev Puri also circulated with a false claim that Puri had ordered additional troop deployments in Delhi at the request of Prime Minister Narendra Modi. The PIB described the video as AI-generated and urged people to rely on official sources.
The pattern is important because these videos do not need to convince everyone. They only need to reach enough people before a correction catches up. A short clip can be removed or debunked later, but screenshots, reposts, and downloaded copies can continue circulating after the original post disappears.
Dursunoglu, Faculty Specialist in Computer Science at Western Michigan University, said the most dangerous AI-generated videos are not always the technically most advanced. The real danger comes when a believable story reaches people at the right emotional moment and gets shared before anyone checks it.
'The most dangerous AI-generated videos going viral right now are not necessarily the most technically sophisticated ones. They are the ones that arrive with a believable story, reach people at the right emotional moment, and are shared faster than anyone verifies them.'
That dynamic was also visible in the United States. In February, Lead Stories fact-checked an AI-generated campaign video falsely presented as a video published by Virginia Governor Abigail Spanberger. The clip used two still photographs from 2023 to create an artificial video of Spanberger appearing to make statements about the Constitution and gun rights.
The video had been posted by the Loudoun County Republican Committee rather than originating from Spanberger.
Another example emerged in the 2026 US election campaign. Reuters reported that an AI-generated advertisement from the National Republican Senatorial Committee depicted Texas politician James Talarico appearing to deliver remarks that were assembled from posts made years earlier.
The footage included a small AI-generated label, but the manipulation was sufficiently realistic to demonstrate how political advertising can use synthetic video to put words into a candidate's mouth.
Not every suspicious video is necessarily AI-generated, either. That distinction matters. In March, social media users claimed footage of Israeli Prime Minister Benjamin Netanyahu showed an extra finger and therefore proved the footage was AI-generated and that Netanyahu was dead.
PolitiFact found that the apparent sixth finger was actually a visual effect caused by the positioning and lighting of the hand. The full footage contained no evidence that the briefing had been generated or manipulated.
That case demonstrates why simply looking for familiar AI 'tells' can produce the wrong answer. Dursunoglu said modern systems are rapidly removing many of the obvious flaws that viewers were once taught to identify.
'One of the biggest misconceptions is that ordinary viewers can reliably identify deepfakes by looking for strange eyes, unnatural blinking, distorted hands, or poor lip synchronization. Those clues can still appear, but modern generative systems are rapidly eliminating many of the obvious artifacts people were taught to look for.'
The Clues Are Changing, So Verification Matters More
Sergey Rodin, General Producer at Lava Media, recommends looking beyond the face because background details can expose problems that are almost invisible during a normal viewing.
'Watch the background, not the subject. Neural networks lose track of complex backgrounds. That's where consistency breaks down. Chair legs swap places between shots. Textures shift. An object quietly disappears from the frame. One person partially merges into another.'
Rodin also recommends watching for missing micro-expressions. A real person makes countless tiny movements around the eyes and mouth without consciously performing them. Generated faces can sometimes reproduce the main expression while failing to reproduce those smaller movements.
'Generated faces don't. The result reads as plastic, flat, slightly glossy, technically clean, but emotionally empty.'
The third clue is physics. Smoke, water, fabric and dust can behave strangely because they require a model to maintain complex movement over time. Rodin points to aircraft and helicopters as useful examples because an apparently realistic shot can still become suspicious when the movement does not match how the object behaves in the real world.
'Worth noting that all three of these are moving targets. The tells that worked six months ago are already weaker today. But right now, these are what still surface generated video most reliably.'
The changing nature of those clues is why verification should start with the source rather than the pixels.
Dursunoglu recommends asking a different question: not whether the video looks fake, but whether its origin can be established.
'That means we need to change the question from "Does this video look fake?" to "Can I establish where this video came from?"'
The first step is to find the earliest available version of the footage. If a video supposedly shows a politician making a major announcement, search the politician's official accounts, government websites and full-length recordings.
If the event supposedly happened at a press conference, look for coverage from established news organisations that were present or have access to the original feed.
The second step is to search for the complete recording. A short clip can remove the context surrounding a statement and make an ordinary comment appear to mean something entirely different.
Comparing a viral segment with the full speech can reveal whether audio has been spliced, whether the speaker's words were rearranged or whether the clip has been digitally manipulated.
The third step is to look for independent confirmation. If a supposedly explosive statement appears only in a viral social media clip, that is a warning sign. Major claims involving heads of government, military action, disasters or financial announcements should normally leave a trail across multiple credible sources.
The fourth step is to examine the file's provenance when the original media is available. C2PA Content Credentials can provide information about where supported media came from and how it has been edited. The C2PA standard is designed to attach verifiable provenance information to digital content, although the absence of credentials does not prove that a video is fake.
Microsoft, for example, uses Content Credentials for supported Azure text-to-speech avatar videos, with information about the video's origin and creation history protected by a cryptographic signature.
OpenAI also uses C2PA metadata and SynthID signals for supported AI-generated images and provides a public tool for checking those signals. The company cautions that a missing signal does not necessarily mean content was not generated by its tools because metadata can be stripped or degraded.
That limitation is crucial. A provenance check can provide evidence, but it is not a universal truth machine.
Dursunoglu said authentication increasingly requires multiple signals, including audio-visual synchronisation, metadata, editing traces, compression patterns, contextual inconsistencies and comparisons with trusted original material.
'No single detector should be treated as a universal truth machine.'
The same principle applies to automated AI detectors. Research published in July 2026 found that traditional artifact-focused approaches are becoming less reliable as generated video improves, with researchers arguing for verification based on whether the people, events, and physical processes shown in a video are consistent with reality.
For ordinary viewers, the safest approach is therefore surprisingly simple. Pause before sharing. Find the original source. Search for the full video. Compare the claim with independent reporting. Examine background details and physical movement. Check available provenance information. Treat familiar faces and voices as clues, not proof.
The financial angle deserves particular caution. Deepfake videos are increasingly being used to make celebrities and public figures appear to endorse investments, financial products or other schemes.
A recent case in Ireland involved a fabricated video falsely showing Taoiseach Micheál Martin promoting a financial product and promising large returns. Martin publicly warned that the video was fake.
Dursunoglu warned that emotional pressure is part of what makes these scams effective.
'The most effective deepfake may be the one that confirms something the viewer already expects or wants to believe.'
That may ultimately be the most important lesson. The question is no longer whether a video looks realistic. AI can increasingly make it look realistic. The question is whether the footage can be traced to a credible source and supported by evidence outside the clip itself.
As Dursunoglu put it, 'We are entering an era in which "seeing is believing" is no longer an adequate security model.' The practical replacement is provenance, corroboration, and careful verification before a sensational video is allowed to become another viral claim.
Frequently Asked Questions
- What are AI-generated videos?AI-generated videos are clips created using artificial intelligence that can mimic real people and events.
- Why are AI-generated videos a concern?They can be used to spread misinformation by making fake events or statements appear real.
- How can you identify a deepfake video?Look for inconsistencies in background details, missing micro-expressions, and verify the video's source and context.
- What should you do before sharing a suspicious video?Verify its authenticity by checking the source, searching for the full video, and comparing it with independent reports.
