AI glossary
Deepfake
A deepfake is synthetic media (video, image or audio) in which a real person’s face, voice or body is generated or altered with deep learning so that they appear to say or do something they never said or did. The term blends “deep learning” and “fake.”
The word entered mainstream awareness in late 2017 when a Reddit user called “deepfakes” posted face-swapped videos made with open-source deep learning tools. Since then, the technology has shifted from a novelty to a tool with serious implications for AI ethics and information integrity.
What makes media a deepfake
At its core, a deepfake is not just a simple cut-and-paste edit. It is content created or modified by Generative AI models that have learned the specific visual or vocal patterns of a target person. While early examples focused heavily on face-swapping, the deepfake meaning now extends to voice cloning, body movement synthesis, and full-media generation.
These systems rely on Computer vision and Image recognition principles to map features like facial landmarks, skin texture, and lighting conditions. The result can range from a subtle change in lip movement to a completely generated person who never existed.
How deepfakes are made
Understanding the mechanics helps explain why detection is becoming harder. Several distinct architectures drive modern synthetic media generation.
Autoencoders and face swaps
The classic face-swap technique uses autoencoders. The model includes one shared encoder that learns a compact representation of faces, and two separate decoders trained on Person A and Person B respectively. When the encoder processes a frame of Person A, passing that data through Person B’s decoder produces an image of Person B wearing Person A’s expression and head pose.
Generative Adversarial Networks (GANs)
Generative adversarial network (GAN) architectures pit a generator against a discriminator. The generator creates synthetic media, while the discriminator evaluates it against real data. This adversarial process was widely used to create photorealistic faces of people who do not exist.
Diffusion models
Diffusion models generate images and video by progressively removing noise from random data until a coherent result emerges. They now produce many realistic synthetic images and videos.
Voice and lip-sync synthesis
Voice cloning uses text-to-speech models trained or conditioned on samples of a person’s voice. Lip-sync models then re-animate a person’s mouth to match the new audio, ensuring the visual movements align with the generated speech.
How to spot a deepfake
Visual and audio inspection remains the first line of defense, though each cue has limitations. Newer generators are rapidly removing many of these artifacts, making manual inspection less reliable over time.
Visual cues
- Unnatural blinking patterns or eye reflections that do not match between the two eyes.
- Lighting and shadows on the face that do not match the surrounding scene or light source.
- Blurring or warping at the edges of the face, hairline, ears, glasses, or teeth.
- Lip movement that appears out of sync with the audio track.
Audio cues
- A flat or oddly paced voice lacking natural rhythm.
- Missing breaths or unnatural emphasis on certain syllables.
- Robotic or overly smooth tonal quality compared to the person’s natural voice.
How to verify media
Because technical artifacts are fading, verification requires a combination of source checking and technical tools.
Source and provenance
Check who published the content first, on which account, and whether reputable outlets have picked up the original version. Run keyframes through a reverse image search to find earlier or original versions of the media.
Digital credentials and watermarks
Content provenance standards are emerging to solve this problem. The C2PA standard (“Content Credentials”) attaches cryptographically signed metadata that records how a piece of media was created and edited. Additionally, some generators embed invisible watermarks in their output; Google DeepMind’s SynthID is one example of this technology.
Detection classifiers
Automated deepfake detection classifiers exist, but they are imperfect. They can fail on content made with newer generation methods, especially if the media has been compressed or re-encoded. Use them as a signal, not proof.
Human verification
For a suspicious voice call asking for money or credentials, hang up and call the person back on a number you already know. Families and companies can agree on a verification question or code word to confirm identity.
Legitimate uses and harms
Deepfakes are a neutral technology. Their impact depends on context and consent.
Legitimate uses
- Film and TV visual effects, such as de-aging actors or restoring footage.
- Dubbing and localization with matching lip movement for international releases.
- Restoring the voice of someone who has lost it due to illness.
- Satire and art created with consent.
Harms and risks
- Non-consensual intimate imagery, often used for harassment or blackmail.
- Fraud, including voice-clone scams that impersonate relatives or company executives to request money transfers.
- Political disinformation designed to sway public opinion.
- Attempts to bypass face or voice-based identity verification systems.
Deepfakes and the law
Legal frameworks are evolving to address synthetic media.
- United States: the TAKE IT DOWN Act was signed into law in May 2025. It criminalizes publishing non-consensual intimate imagery, including AI-generated deepfakes, and requires covered platforms to remove such content within 48 hours of a valid request.
- European Union: the EU AI Act (Article 50) requires deployers of AI systems that generate or manipulate image, audio or video content constituting a deepfake to disclose that the content has been artificially generated or manipulated.
FAQ
What is a deepfake?
A deepfake is synthetic media (video, image or audio) in which a real person’s face, voice or body is generated or altered with deep learning so that they appear to say or do something they never said or did.
How are deepfakes detected?
Detection relies on spotting artifacts like unnatural blinking, mismatched lighting, or lip-sync errors. Automated classifiers and provenance standards like C2PA also help verify authenticity, though no method is foolproof.
What is the difference between deepfake and AI-generated video?
All deepfakes are AI-generated, but not all AI-generated video is a deepfake. A deepfake specifically involves manipulating or generating media to depict a real person (or their likeness) in a fabricated context.
Are deepfakes illegal?
Legality depends on jurisdiction and context. The EU AI Act requires disclosure of manipulated media, while the US TAKE IT DOWN Act targets non-consensual intimate imagery. Using a person’s likeness for satire or commentary may be protected speech, depending on local laws.