Definition:
A deepfake is an image, video or audio recording generated or altered using artificial intelligence to plausibly represent people or situations that do not correspond to a recording of a real event. It can make someone appear to say or do something they never said or did, or construct a fictional identity.
The term combines deep learning and fake. It describes synthetic or manipulated content, not a single technique. Its use may be creative and consensual, or serve to deceive and impersonate people.
Table of contents
History and evolution of deepfakes
Photographs, films and recordings were being manipulated long before artificial intelligence. Deep learning expanded the ability to learn visual and audio characteristics from examples and use them to generate new representations.
During the second half of the 2010s, face-swapping tools and manipulated videos shared on forums and social media became popular. Their uses included celebrity montages and the creation of intimate content without consent.
The evolution of generative AI extended these capabilities to voice cloning, lip synchronisation and image and video generation. There is no single architecture behind all deepfakes: autoencoders, generative adversarial networks and diffusion models have been used, among other techniques.
How deepfakes work
Systems learn patterns from data and use them to synthesise or transform content. The result depends on the model, reference material and subsequent processing; not all systems need the same amount of data or specific training for each person.
In a generative adversarial network, or GAN, a generator produces samples and a discriminator learns to distinguish them from real samples during training. The discriminator is not an editor that directly corrects every imperfection in the video. Other systems use different processes and do not include this pair of components.
Common transformations include face swapping, recreating expressions, cloning a voice and synchronising lip movements with audio. They can be combined, but changing a face does not imply reproducing that person’s voice as well.
Deepfakes fall within the field of generative AI, although not all AI-generated content is a deepfake. A synthetic graphic or generic artificial voice does not necessarily involve impersonation or a representation that could be mistaken for a real recording.
Applications of deepfakes
Identity synthesis and modification techniques can be used in different fields. How they should be regarded depends on the content, consent and how they are presented to the public:
- Entertainment and film: altering performers’ appearances, recreations and audiovisual dubbing. Not every visual effect is a deepfake.
- Video games: reproducing appearances or voices with permission. Creating a completely stylised character does not, by itself, amount to generating a deepfake.
- Advertising and marketing: adapting material to other languages or markets and authorised use of images or voices. A recreation should not be confused with a genuine product testimonial.
- Education and training: audiovisual reconstructions and simulations, identified as such so that they are not treated as documentary evidence.
- Customer service: authorised reproduction of a vocal identity in conversational interfaces. A synthetic voice that does not imitate a specific person is not necessarily a deepfake.
- Research: creating manipulated samples to study perception or evaluate detectors. Not all synthetic datasets fall into this category.
The same capabilities can be used to spread disinformation, harass people or commit fraud. The technique alone does not determine whether a use is legitimate.
How to identify deepfakes
No single visual sign or tool can resolve every case. Verification can combine different sources of evidence:
- Origin and context: locate the initial publication, review who is sharing it and cross-check the alleged event with other independent sources.
- Metadata and provenance: examine available information about creation and editing. Metadata can be lost or modified; their absence does not prove that a file is fake.
- Image and movement: look for inconsistencies in faces, lighting, edges or movements. Compression and other legitimate processes can also produce anomalies.
- Audio and synchronisation: review cuts, voice changes or lip-sync mismatches. Dubbing and transmission problems may explain some of them.
- Automatic detectors: treat their results as indications, considering false positives, false negatives and differences between the material used to train them and the material they analyse.
- Combined analysis: bring together technical results, original sources and specialist review when the importance of the case warrants it.
C2PA-based Content Credentials allow signed information about a file’s provenance and modifications to be associated with it. They provide traceability, but do not, by themselves, certify that what is depicted actually happened. Nor does the absence of credentials demonstrate manipulation.
Unusual blinking or an unnatural voice may warrant further review, but are not conclusive evidence. Similarly, a convincing-looking video does not establish its authenticity.
Impact of deepfakes on society and culture
Deepfakes can affect reputation, privacy and trust in audiovisual evidence. Harmful uses include spreading fabricated statements, impersonation in calls or video calls, and creating intimate images without consent.
In fraud, a recognisable voice or face may be used to lend credibility to a request for money or information. An appearance of familiarity does not establish who is actually controlling the communication.
There is also a risk that an authentic recording may be discredited by claiming it was AI-generated. Assessment should therefore assume neither that all content is fake nor that every convincing image is genuine.
The consent of the people represented, identification of synthetic content and the ability to check its provenance are relevant considerations when assessing these uses. Legal obligations depend on the territory and context; the mere existence of a tool does not authorise every use of another person’s image or voice.