What Is a Deepfake and How Is It Created?

Defining the Term

A deepfake is synthetic media, most commonly video, image, or audio, generated or manipulated using AI techniques to depict a real, identifiable person saying or doing something they did not actually say or do. The term itself is a combination of “deep learning,” the category of AI techniques generally underlying the technology, and “fake,” referring to the fabricated nature of the resulting content. While the term is sometimes used loosely to describe any AI-manipulated media, it more precisely refers to content specifically designed to realistically depict a real person in a fabricated scenario.

This distinguishes deepfakes from other categories of AI-generated content, like an entirely fictional AI-generated image or a stylized virtual avatar that isn’t designed to closely resemble a specific real individual.

The Core Techniques Behind Deepfake Creation

Deepfake creation generally relies on training an AI model using existing images, video, or audio of the target person, the more reference material available, and the higher its quality, the more convincing the resulting output tends to be. Several distinct techniques fall under the deepfake umbrella. Face-swapping involves replacing one person’s face in existing footage with another person’s likeness. Facial reenactment involves mapping a source person’s expressions and movements onto a target person’s face without swapping identity entirely. Voice cloning uses similar AI training approaches to replicate a specific person’s vocal characteristics and speech patterns, allowing new audio to be generated that sounds like that person speaking words they never actually said.

These techniques have become substantially more accessible over time, as tools and applications have packaged the underlying AI models into more user-friendly interfaces, lowering the technical expertise historically required to produce a convincing result.

Legitimate Applications Alongside the Well-Known Risks

It’s worth noting that the same underlying technology isn’t exclusively used for harmful purposes. Film and television productions have used related AI techniques for legitimate creative purposes, including digital de-aging of actors and adjusting an actor’s lip movements to match dubbed dialogue in a different language for international releases. These applications generally involve the consent and participation of the person being depicted, a key distinction from the unauthorized and often deceptive uses of deepfake technology that have driven the most significant public and legal concern, covered in more depth elsewhere in this topic.

Bottom Line

A deepfake is AI-generated or AI-manipulated synthetic media that realistically depicts a real, identifiable person saying or doing something fabricated, created by training deep learning models on reference footage or audio of that person, with techniques ranging from face-swapping to voice cloning that have become increasingly accessible even as legitimate, consent-based applications of similar technology also exist.

Go deeper

• Can ai detect deepfake voice calls in real time during a phone call?