what’s a deepfake?

the word was coined in late 2017 by a Reddit user with the handle deepfakes, who ran a subreddit sharing face-swap videos made with an open-source autoencoder. the original technical meaning was narrow: face replacement in video. usage has since broadened to cover any AI-synthesized depiction of a real person.

researchers split the category three ways. face swap replaces one person's face with another's in otherwise authentic footage. lip sync or reenactment keeps the original face but reshapes the mouth and expression to match new audio, which makes it the hardest variant to spot by eye. full synthesis generates an entirely fabricated person or scene, including voice clones and talking-head avatars like those used in customer-service videos.

volumes climbed sharply through 2024. the identity-verification vendor Sumsub reported a 4× year-over-year increase in deepfakes detected globally during the year of major democratic elections, and a 303% spike in the U.S. ahead of November. despite the worst fears, documented political deepfakes that meaningfully changed an election outcome remain rare. the dominant real-world harm is non-consensual intimate imagery and financial fraud (especially voice-clone scam calls), not propaganda.

the most common misconception, including in news copy, is that a deepfake must involve video. audio-only voice clones used in scam calls are deepfakes too, and as of 2024 they are the most economically damaging variant.

deepfake detection is harder than generic AI-image detection because the synthetic part of the content is often spatially or temporally local. a 30-second face swap inside an otherwise-real five-minute interview will fool a detector that averages a global score across frames. amige. handles this by trusting the loudest deepfake signal from either detector. that's also why our headline number for a flagged deepfake can be higher than the unweighted detector average.

bigger picture: in-the-wild detector accuracy on deepfakes is materially worse than in academic benchmarks. open-source detectors drop roughly 45 to 50% in AUC when moved from clean lab evaluation sets to actual social-media content (Deepfake-Eval-2024). commercial vendors report higher numbers; treat the lab figures as ceilings, not promises.