A deepfake is synthetic media, usually a video, audio clip or image, in which a real person’s face, voice or actions are convincingly altered or fabricated with artificial intelligence. The word combines deep learning and fake. A deepfake may swap one person’s face onto another’s body, clone a voice from a short recording, change the words a speaker appears to say or generate a complete video of an event that never happened. The technology is closely related to generative AI and to tools for video generation, speech generation and image generation. It has legitimate creative uses, but it is best known for fraud, harassment and disinformation.
History of deepfakes
Manipulating images is old, and film studios have long used visual effects to alter faces. Academic work on face reenactment, such as Video Rewrite in 1997 and Face2Face in 2016, showed that computers could change a speaker’s mouth to match different words. The modern term appeared in late 2017, when a user on Reddit with the name deepfakes shared videos that put celebrities’ faces onto other actors’ bodies using autoencoders. Open-source tools followed, and the volume of fake content grew. In 2018 and 2019, generative adversarial networks produced fake faces that looked real, a widely shared video of a politician appeared with a manipulated speech and a fake video of a public figure warning about misuse was made for awareness. Voice cloning came into use for fraud, and in 2024 a company in Hong Kong reportedly lost about 25 million dollars after staff joined a video call with deepfake versions of executives. By the mid 2020s, text-prompted video and image models made convincing fakes available to anyone.
How deepfakes work
Several techniques are used.
- Face swapping with autoencoders: two networks share an encoder that learns facial features and use different decoders for different people. Encoding person A and decoding with B’s decoder produces B’s face with A’s expression. See variational autoencoder.
- GAN-based synthesis: a generator creates faces and a discriminator tries to tell them from real ones, improving realism through competition.
- Face reenactment and lip sync: models transfer expressions and mouth shapes from a driving video or audio to a target face.
- Diffusion and video models: modern diffusion models create images and clips of specific people from a few reference photos.
- Voice cloning: text-to-speech systems imitate a person’s voice from seconds of audio.
Earlier deepfakes needed many images of the target and a powerful computer. Today, a handful of photos and an online tool can be enough for a convincing result.
Legitimate uses
The same methods support film dubbing in which actors’ mouths match translated speech, de-aging of actors, restoring voices for people who lost speech, privacy-preserving anonymization of faces in documentaries, educational recreations of historical figures with clear labels, satire clearly marked as parody and training data for detection systems. These uses depend on consent, transparency and context.
Harms and risks
- Non-consensual intimate imagery: the largest share of deepfake content found online has been sexual material that places people, mostly women, in explicit scenes without consent. This causes serious harm to victims.
- Fraud and scams: cloned voices and faces are used to trick relatives, employees and banks into sending money or revealing information.
- Disinformation and politics: fabricated clips of leaders or candidates can mislead voters, and the existence of fakes allows real evidence to be dismissed as fake, a problem called the liar’s dividend.
- Harassment and reputation harm: deepfakes are used to bully, extort and defame.
- Erosion of trust: if any recording might be fake, public trust in audio and video evidence declines.
Detection
Detecting deepfakes is a race between creators and defenders. Detectors look for artifacts such as unnatural blinking, inconsistent lighting, odd teeth and hands, mismatched audio and lip movement, or traces left by generation methods in the pixels or sound. Machine learning classifiers are trained on known fakes, but they often fail on new techniques, so no tool is fully reliable. A complementary approach is provenance. Standards such as C2PA attach signed information to media at creation, recording the device or software used and any edits, and some generators add invisible watermarks. Platforms combine automated checks, user reports and human review.
Laws and policy
Laws are developing quickly. Many countries and US states have criminalized non-consensual sexual deepfakes, and several have rules against deceptive synthetic media in election campaigns. The European Union AI Act requires labeling of deepfakes, and China requires labeling and registration of synthetic content. In the United States, the TAKE IT DOWN Act of 2025 requires platforms to remove non-consensual intimate images, including AI-generated ones, after notice. Platforms such as YouTube, Meta and TikTok have policies that require disclosure of realistic synthetic content. Rights of publicity and defamation law may also apply.
How to protect yourself
Be cautious with urgent requests for money, even when the voice or face seems familiar, and verify through a second channel. Agree on a family code word. Check the source of surprising clips and look for coverage from trusted outlets. Limit public posting of high-quality voice and face material where practical, and use privacy settings. If you are targeted, save evidence, report to the platform, and contact local authorities or support organizations. Organizations should use callback procedures for payments and train staff on voice and video fraud.
A worked example
An accountant receives a video call from someone who looks and sounds like the finance director, requesting an urgent transfer to a new account. The accountant notices that the request breaks company policy and hangs up. She calls the director on his known number, learns the call was fake and reports the incident. The company adds a rule that any payment to a new account needs approval through a second channel, and it begins training staff on deepfake fraud.
FAQs
Are all deepfakes illegal? No. Many are legal when consensual and labeled, but creating sexual deepfakes without consent, fraud and election deception are illegal in many places.
Can deepfakes be detected reliably? Not always. Detection tools help, but provenance data and careful verification of the source are also needed.
What should I do if I find a deepfake of me? Save evidence, report it to the platform, ask for removal and seek legal or support help where needed.
More in Generative AI
Key facts
- The word deepfake comes from deep learning and fake, and first spread in 2017 on Reddit, where a user shared face-swapped videos.1
- Most detected deepfake videos online have been non-consensual sexual content, according to analyses by security researchers.2
- Provenance standards such as C2PA content credentials help show where media came from.3
Related concepts
Quick checkWhat is a deepfake?Show answer
It is media created or altered by AI to make a real person appear to say or do something they did not, often with realistic face swaps, voice cloning or full video synthesis.
Sources
- Chesney and Citron, Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security, https://scholarship.law.bu.edu/faculty_scholarship/640/
- Sensity AI and Deeptrace, The State of Deepfakes, https://regmedia.co.uk/2019/10/08/deepfake_report.pdf
- Coalition for Content Provenance and Authenticity, C2PA, https://c2pa.org/
Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue
