A deepfake is a video in which a person's face, voice or entire appearance has been generated or replaced using artificial intelligence, making them say or do things that never happened. Until a few years ago the results were clumsy and easy to expose. Today some are almost indistinguishable from reality at first glance. This does not mean they cannot be recognised, but that the approach has to change: less trust in the eye, more attention to context.
First rule: stop before you share
A deepfake succeeds when it triggers a strong, immediate emotion: outrage, fear, amusement. That is exactly the moment people share without thinking. If a video seems sensational, scandalous or too perfect a confirmation of what you already thought about someone, that is your cue to slow down.
Visual signs to look for
Generators keep improving and some typical flaws are disappearing. But there are still places where errors often cluster:
- The outline of the face: unnatural blending between face and hair, neck or ears, especially when the person turns their head.
- The mouth and teeth: teeth that look like a single block, a poorly defined tongue, lip movements not perfectly in sync with the words.
- The eyes: irregular blinking, a fixed stare, different reflections in each eye.
- The skin: a surface that is too smooth, almost plastic, or that changes texture from frame to frame.
- Hands and objects: fingers merging together, objects appearing or warping as they pass in front of the face.
- The lighting: shadows on the face that do not match the light in the scene.
A simple trick: play the video at reduced speed and pause during fast movements. Many flaws that are invisible at normal speed become obvious frame by frame.
Audio signs
When the voice is synthetic too, pay attention to:
- flat intonation or emphasis in the wrong places;
- no breaths, natural pauses or hesitations;
- audio that is too clean for the setting shown, or background noise that never changes;
- unusual pronunciation of proper names or foreign words.
Context checks, the most reliable
Because visual flaws are becoming rarer, the most solid checks concern everything around the video.
Who published it first?
Look for the original source. An authentic video of a public statement is usually found on the person's official channels or those of recognised news outlets. If it only circulates on anonymous, recently created accounts or ones specialising in provocative content, that is a red flag.
Where and when was it supposedly filmed?
Compare the location, clothing and date with independent sources. If the person was demonstrably elsewhere at the time, or the event is not reported anywhere else, the likelihood of a fake rises sharply.
Are there different versions?
Take one or two significant frames and run a reverse image search. You may find that the original video exists, but with different audio or a different face.
Is it consistent with the person?
Are the language, manner of speaking and views expressed compatible with what the person has said publicly so far? A radical change of position announced in an isolated video deserves caution.
Automated tools: useful, not decisive
There are online services that claim to determine whether a video was generated with AI. They can offer clues, but they have important limits: they can be wrong in both directions and struggle with compressed videos like those circulating on social media. Use them as one element of your assessment, never as a final verdict.
Some platforms and some makers of generation tools are introducing labels and information about the origin of content. When present they are useful, but their absence does not prove a video is authentic.
If the video is about you
If you discover a deepfake depicting you, do not just deny it. Save the address and take screenshots of the page, flag the content to the platform using the category for manipulated media or impersonation, and post a denial on your official channels, short and without amplifying the video.
In short
- A video that triggers strong, immediate emotions deserves checking before it is shared.
- Look at the face outline, mouth, eyes, skin and lighting, including in slow motion.
- The most reliable checks concern context: original source, place, date and consistency.
- A reverse image search on frames can reveal the original video.
- Automated tools help, but they do not replace reasoning.