What a deepfake is

A deepfake is a piece of media, such as a video, image, or audio clip, that has been created or altered using artificial intelligence to look convincingly real. The most familiar type shows a real person’s face or voice placed into a situation that never happened, making it appear that they said or did something they did not. The word itself is a blend of “deep learning,” a branch of artificial intelligence, and the word “fake.”

The technology first drew wide public attention around 2017, when methods for swapping faces in video became accessible to ordinary users. Since then it has expanded well beyond video to include synthetic still images and cloned voices. What unites all of these is that a computer has generated or modified the content rather than a camera or microphone simply recording reality.

It is worth being clear that the technology underneath is neutral. The same methods are used for harmless and creative purposes, such as visual effects in films, satire, dubbing films into other languages, and art. The term “deepfake” tends to be used when such media is produced in a misleading or deceptive way, which is why it carries a negative tone in the news.

How deepfakes are made

Deepfakes are built using machine learning, a type of artificial intelligence in which a system improves at a task by studying large amounts of example data rather than following fixed, hand-written rules. To create a convincing fake of a person, these systems are typically trained on many images, video frames, or recordings of that person. From this material the model learns the patterns of how a face moves, how it is lit from different angles, or how a voice sounds across many words and tones.

One influential approach uses a setup in which two artificial neural networks are pitted against each other, a design known as a generative adversarial network. In this arrangement one network, the generator, tries to produce fake images, while a second network, the discriminator, tries to tell the fakes apart from real examples. The two improve in competition: as the discriminator gets better at catching fakes, the generator is pushed to make them more convincing. Over many rounds this contest drives the generated results toward being increasingly realistic.

Other modern techniques generate images and video in different ways, but the underlying principle is similar. The system learns the statistical patterns of real media so well that it can produce brand-new content that shares those patterns. This is also how tools that create entirely imaginary faces or scenes from a text description work, and the line between “editing a real clip” and “generating a new one” has blurred as the technology has advanced.

From faces to voices

Audio deepfakes, often called voice cloning, apply the same idea to sound. Given enough samples of someone speaking, a model can learn the characteristics of their voice and then generate new speech in that voice saying words they never recorded. Combined with video, this makes it possible to produce clips that look and sound like a real person, which is part of why the topic has become a serious concern.

Why deepfakes can be so convincing

Two forces have made deepfakes steadily more believable. The first is the sheer volume of training material available online. Public photos and videos give these systems plenty to learn from, especially for well-known figures who appear in large amounts of footage. The second is that the tools have become faster, cheaper, and easier to use, so producing a fake no longer requires specialist skills or expensive equipment.

As a result, the quality of fakes has risen quickly, and the obvious glitches that once gave them away are becoming less common. This is an important point for readers: advice on spotting deepfakes is useful, but it is not foolproof, and it should be combined with healthy scepticism about where a clip came from.

How to spot a deepfake

While detection is getting harder, many fakes still contain clues, particularly lower-quality ones. When you watch a suspicious clip, it can help to look closely for the following.

  • Unnatural faces and movement. Watch for strange blinking patterns, skin that looks too smooth or waxy, or facial expressions that do not quite match the emotion of the words.
  • Edges and blending. Look at the boundaries of the face, hair, and glasses, and at the neck and ears, where fakes sometimes show blurring, flickering, or mismatched skin tones.
  • Lighting and shadows. Inconsistent lighting, odd reflections in the eyes or on glasses, and shadows that fall the wrong way can reveal manipulation.
  • Audio mismatches. Check whether the lip movements line up with the sound, and listen for a voice that sounds flat, oddly paced, or lacking natural background noise.
  • Glitches at transitions. Pausing and stepping through a video frame by frame can expose warping when the head turns or something passes in front of the face.

Beyond these visual and audio cues, the most reliable defence is context. Ask where the clip originated, whether reputable sources are also reporting the same thing, and whether the content is designed to provoke a strong emotional reaction, which is a common feature of deceptive media. A surprising or inflammatory clip from an unknown account deserves extra caution.

How technology fights back

Researchers are actively building tools to detect synthetic media automatically, often by training systems to recognise subtle traces left behind by the generation process that human eyes cannot see. However, detection is a genuinely difficult and unsolved problem. Studies have repeatedly found that detectors trained on one method often fail to catch fakes produced by newer or different methods, so detection tends to lag behind the latest generation techniques.

Because of this, other approaches are being developed alongside detection. These include methods for marking or watermarking AI-generated content so it can be identified, and systems for recording the origin and editing history of a piece of media so its authenticity can be checked. No single solution is complete, and most experts expect a combination of technical tools, platform policies, and public awareness to be needed.

The risks and why they matter

Deepfakes raise several real concerns. The most discussed is misinformation: fabricated clips of public figures could spread false statements quickly, especially during sensitive moments such as elections. There are also fraud risks, where a cloned voice or video might be used to impersonate someone and trick a person or business into transferring money or revealing information. Deepfakes have additionally been used to create harmful fake images of individuals without their consent, which is a serious abuse that many places are moving to outlaw.

A subtler danger is the erosion of trust itself. If people come to believe that any video or recording could be fake, genuine evidence can be dismissed as fabricated, a problem sometimes described as the “liar’s dividend.” Protecting a shared sense of what is real therefore matters beyond any single faked clip.

Staying grounded

For most people, the practical response to deepfakes is not fear but informed caution. Understanding that convincing media can be generated by artificial intelligence changes how you read a sensational clip: instead of taking it at face value, you can pause, consider the source, and look for confirmation from trusted outlets before believing or sharing it. As the technology continues to improve, that habit of verifying before trusting is likely to become one of the most valuable skills for navigating life online.