How to tell whether a video is a deepfake
Deepfakes are turning up in election ads, scam promotions and harassment campaigns, and the old visual giveaways are disappearing. What still helps, and what to do if you are targeted.
After the US midterm elections, a video spread online that appeared to show Donald Trump destroying an office. It had been generated with artificial intelligence, as the fact-checking site Newschecker later reported. By then it had already been shared widely.
Cases like this are now routine. A few years ago, spotting a deepfake was largely a matter of looking for obvious mistakes, such as extra fingers, flickering ears or faces that never blinked. The tools have improved to the point where many of those mistakes no longer appear, and the people making fakes have learned that speed matters more than perfection. A video only has to be believable for the hour or two it takes to go viral.
Where deepfakes are appearing
Election campaigns have become one of the most visible battlegrounds. In the Australian state of Victoria, the major parties have traded accusations over AI-generated videos of political leaders used in election advertising. In Nigeria, the electoral commission has warned the media against publishing fake results and AI deepfakes ahead of the 2027 elections.
Fraudsters use the same technology to sell products. A Facebook advertisement was recently found using deepfakes of public figures to promote an unverified “eyesight cure”. Far more damaging is the use of deepfakes against private individuals. A study reported this month found that sexually explicit deepfakes of ordinary people, rather than celebrities, are becoming more common, and Apple and Google are facing scrutiny over more than 400 apps and websites linked to AI deepfakes.
Courts have started to respond. In India, the Delhi High Court restrained the circulation of AI deepfakes of the actor Kriti Sanon, one of several recent rulings protecting public figures’ likenesses. Security teams have a related worry: deepfaked faces and voices are being used to get past account recovery checks that rely on video or voice verification.
What the eyes can still catch
Lower-quality fakes still give themselves away, and it is worth knowing where to look. Hands, teeth, ears and jewelry are often rendered imperfectly. The edges where a face meets hair, glasses or the background can blur or shimmer. Light on a face may not match the room, and shadows can fall in the wrong direction. Lip movements are sometimes slightly out of step with speech, and any text in the frame, such as signs or labels, may be garbled. In audio, a flat emotional tone, odd breathing or background noise that cuts in and out can be signs of manipulation.
The difficulty is that the best tools now avoid most of these errors. Finding no flaw is not evidence that a video is genuine.
Checking the context instead
Professional fact-checkers spend less time studying pixels than asking where a piece of content came from. The first question is who posted it originally: a news organization, the person’s own verified account, or an account with no history. The second is whether anyone else is reporting it. If a well-known figure had really said or done something shocking, several established outlets would be covering it, and their silence is itself informative.
A reverse image search on a screenshot can turn up earlier versions of the footage or the original it was built from. Some cameras, editing programs and platforms now attach Content Credentials, a record of how an image was made and changed. Their absence proves nothing, but where they are present they can help confirm that an image is authentic. Detection software exists, though none of it is reliable enough to settle the question on its own.
It also helps to notice how a video makes you feel. Fakes made to deceive are usually designed to provoke anger or fear, because those emotions make people share before they check.
If someone makes a deepfake of you
People who discover a deepfake of themselves are generally advised not to engage with whoever made it, and never to pay if they are being threatened. Screenshots, links, usernames and dates should be saved as evidence, and the content reported to the platform hosting it, most of which have reporting options for impersonation and for intimate images shared without consent. For intimate images of adults, the StopNCII.org service can help stop them spreading across participating platforms, and for images of anyone under 18, the US National Center for Missing and Exploited Children runs a removal service called Take It Down. Cases involving threats, extortion or children should be reported to the police.
As generation tools improve, the question people face is shifting from whether they can spot a flaw to whether they have a good reason to believe what they are seeing. For now, the most reliable answer still comes from checking the source rather than the screen. New cases are logged on our deepfakes page.
Sources
Show all 8 sources
Introduction
- Newschecker: AI-Generated Video Goes Viral As Showing Trump Destroying Office After US Midterms Setback
Where deepfakes are appearing
- The Guardian: Scare campaigns: Victoria’s major parties trade blows over AI-generated videos of leaders in election ads
- Nigerian Bulletin: 2027 Elections: INEC Warns Media Against Fake Results, AI Deepfakes
- NewsMeter: Facebook ad uses AI deepfakes of public figures to promote unverified ‘eyesight cure’
- U.S. News & World Report: Sexually Explicit Deepfakes Of Average Citizens Becoming More Common, Study Says
- TechRepublic: Apple and Google Face Scrutiny Over 400 Apps and Websites Linked to AI Deepfakes
- MediaNama: Delhi HC restrains AI deepfakes of Kriti Sanon in interim order
- TechTarget: MFA recovery becomes a deepfake weak link