A deepfake is a piece of synthetic media, most often video or audio, that has been generated or manipulated using deep learning techniques to convincingly depict a real person saying or doing something they did not actually say or do. The term is a portmanteau of deep learning and fake, and originated in late 2017 on a Reddit forum where a user named deepfakes posted face-swapped videos, initially of celebrities, using early generative adversarial network-based face-swapping tools.
Deepfake technology advanced rapidly through the late 2010s and early 2020s, moving from crude, easily detectable face swaps to highly convincing video and, increasingly, real-time voice cloning and full-body synthesis, raising concerns spanning non-consensual pornography, political disinformation, and fraud.
Techniques
Early deepfakes primarily used autoencoder-based face-swapping, training a shared encoder and separate decoders on footage of two people so that one person's expressions could be mapped onto the other's face. GAN-based methods improved realism further by pitting a generator against a discriminator trained to detect fakes. By the mid-2020s, diffusion models had become the dominant technique for high-quality synthetic video and image generation, and dedicated face-swap and lip-sync tools, alongside general-purpose text-to-video systems, made producing a convincing fake video dramatically cheaper and faster than in the technology's early years.
Malicious and legitimate uses
Documented malicious applications include non-consensual intimate imagery, the earliest and still most common reported use case, political disinformation such as fabricated statements by public figures released around elections, and financial fraud, including a widely reported 2024 case in which a finance employee at a multinational firm was deceived by a video call featuring deepfaked executives into wiring millions of dollars. Deepfake technology also has acknowledged legitimate applications, including film dubbing and de-aging effects, accessibility tools that generate a synthetic voice for people who have lost their own, and satire protected as free expression in some jurisdictions.
Detection and countermeasures
Detecting deepfakes has become an active research area, using artifacts such as inconsistent blinking, unnatural facial geometry, or subtle audio-visual mismatches, though detection accuracy has struggled to keep pace with generation quality, particularly for the highest-end models. Watermarking and content-provenance standards have been proposed as complementary defenses that do not rely on spotting generation artifacts. Multiple jurisdictions moved to regulate deepfakes specifically in the mid-2020s, including US state-level laws targeting election-related and non-consensual intimate deepfakes and provisions within the EU AI Act requiring disclosure of AI-generated media, though enforcement across borders remained difficult given how easily such content can be produced and distributed.
Impact
Beyond specific harmful incidents, researchers and journalists have described a broader liar's dividend: as the public becomes aware that convincing fake video and audio are possible, genuine footage can be dismissed as fabricated, undermining trust in authentic evidence generally. This dynamic has been cited as a factor in the growth of dead internet theory and related discourse about the reliability of online media in the generative AI era.