The name combines 'deep learning' and 'fake' and spread from 2017 onwards. Deepfakes range from harmless parody to fraud, harassment and political misinformation. Platforms and some laws now require realistic AI-generated content to be labelled.
How it works
Most deepfakes rely on models that learn what a person looks or sounds like from examples. Classic face-swap tools train an autoencoder on many images of two faces: a shared encoder learns expressions and pose, and separate decoders learn each identity. Feed person A's video through person B's decoder and B's face appears with A's expressions. Generative adversarial networks, where a generator tries to fool a discriminator, pushed realism further, and diffusion models now generate convincing faces and whole scenes from text. Lip-sync models can make an existing video match new audio. Voice deepfakes use voice cloning, which can copy a voice from a short recording. Detection looks for clues such as unnatural blinking, mismatched lighting, odd teeth or ears, and audio artefacts, but these clues fade as models improve. Because of that, the industry is also working on provenance: standards such as C2PA content credentials attach signed information about how a file was made, and some AI tools add invisible watermarks.
Why it matters
Deepfakes affect anyone with a public face or voice. Creators risk having their likeness used in fake endorsements or scams. Businesses face fraud where criminals imitate an executive's voice to request urgent payments. For everyone, they make it harder to trust video as evidence. Knowing how they work helps you set verification rules, protect your brand, label your own AI content correctly and respond fast when a fake of you appears.
Where you will see it
- Film and TV de-ageing actors or completing scenes after an actor is unavailable, with consent.
- Satire and parody videos clearly labelled as fake.
- Security teams training staff to recognise voice and video impersonation scams.
- Consented AI avatars and translated videos for training content.
Limits and common misconceptions
- Used without consent, deepfakes can be illegal or break platform rules, including fraud, defamation and non-consensual intimate imagery.
- Detection tools are unreliable; a 'clean' result does not prove a video is real.
- Not every AI-edited video is a deepfake; the term usually means imitating a real, identifiable person.
How it differs from related terms
- Deepfake vs Voice cloning Voice cloning copies how someone sounds; a deepfake is the broader category of fake media of a real person, which may include a cloned voice.
- Deepfake vs Lip-sync Lip-sync is one technique that matches mouth movement to audio; it becomes a deepfake when used to put words in a real person's mouth.
Frequently asked questions
Are deepfakes illegal?
Making one is not automatically illegal, but many uses are. Fraud, defamation, harassment, election interference and non-consensual intimate deepfakes are covered by laws in many countries, and new rules are being added. The EU AI Act includes transparency duties for deepfakes. Platforms also have their own labelling and removal policies.
How can I tell if a video is a deepfake?
Look for odd blinking, blurry edges around the face, mismatched lighting, strange teeth or hands, and lips slightly out of sync. Check the source, search for the original and look for content credentials. Treat urgent requests for money or secrets with extra suspicion, whatever the video shows.
Do I need to label AI-generated videos?
On major platforms, usually yes when content is realistic and could mislead viewers about real people or events. YouTube, TikTok and Meta all have disclosure options or requirements for realistic synthetic media. Rules differ by platform and country, so check current policies before you publish.
Related terms
- Copilot An AI assistant embedded inside another app that uses what you are working on as context to suggest, draft or complete tasks.
- Deep learning Machine learning that trains neural networks with many stacked layers to learn patterns directly from raw data such as text, pixels or sound.
- Diffusion model A generative model that creates images, video or audio by turning random noise, step by step, into a result that matches the prompt.
- Dubbing (AI) Using AI to translate a video's speech into another language and re-voice it, often in a clone of the original speaker's voice.
- Embedding A list of numbers (a vector) that captures the meaning of text, images or audio, so that similar items end up close together mathematically.