Deepfake meaning: where the word came from and what it covers

This page belongs to the deepfake detection hub and is the etymology companion to the definition of a deepfake, which covers what the thing is and how well it is detected. Here the subject is the word itself: what each dictionary says, what the research papers say, and where the two senses part.

01What does deepfake mean in the dictionaries?

Four dictionary entries were read for this page, and they agree on the parts and disagree only on the date. Merriam-Webster defines a deepfake as "a digital image, video, voice, etc. that has been convincingly generated or altered to misrepresent someone as doing or saying something that was not actually done or said", and in its usage note adds that the deep half "is specifically influenced by deep learning" (Merriam-Webster). It records a transitive verb, to deepfake, and an adjective, deepfaked.

The Oxford Advanced Learner's Dictionary has "a video or audio recording that has been digitally altered in order to present a particular person doing or saying something that they never actually did or said", with the origin given as "early 21st century: from deep (adjective) and fake (noun)" (Oxford Learner's Dictionaries). Dictionary.com defines it as "a fake, digitally manipulated video or audio file produced by using deep learning, an advanced type of machine learning" and dates it as first recorded between 2015 and 2020 (Dictionary.com). The Cambridge Advanced Learner's Dictionary is the narrowest: "a video or sound recording that replaces someone's face or voice with that of someone else, in a way that appears real", with no date (Cambridge Dictionary).

Three things are common to all four. The medium is video or audio, with Merriam-Webster adding still images. The maker is a learned model, named as deep learning in two of the entries. And the effect is misrepresentation of a real person. Only Cambridge restricts the method to replacement of a face or voice; the other three cover anything "generated or altered".

02When was the word first used?

The dictionaries give three different answers, and this page quotes each as printed rather than choosing one. Merriam-Webster: "the first known use of deepfake was in 2018". Dictionary.com: first recorded between 2015 and 2020 (printed as a year range). Oxford Advanced Learner's: "early 21st century". A dictionary's first-use date is the earliest attestation its editors have on file in the sense they define, not a claim about when the technique was invented, which is why the dates can differ between publishers and move as earlier citations are found.

The earliest research record read for this page is the FaceForensics++ paper, submitted to arXiv in January 2019 and published at ICCV 2019 (Rössler et al., arXiv 1901.08971). Its authors write that "the term Deepfakes has widely become a synonym for face replacement based on deep learning, but it is also the name of a specific manipulation method that was spread via online forums", and that "to distinguish these, we denote said method by DeepFakes in the following paper". The paper does not name a person, a forum, an account or a month for the method's first appearance, and no primary source fetched for this page does. Accounts that trace the name to a specific online account are widely repeated, but they rest on press reports, and this site does not state them as fact.

03What did the word mean in the first papers?

In its narrow 2019 sense, DeepFakes with a capital F was one method among four in the FaceForensics++ benchmark, alongside Face2Face, FaceSwap and NeuralTextures. The paper describes it as "based on two autoencoders with a shared encoder that are trained to reconstruct training images of the source and the target face", with public implementations "most notably FakeApp and the faceswap github". The how deepfakes are made page walks through that pipeline step by step.

The introduction of the same paper shows the word's connotation from the start: the method "has shown how computer graphics and visualization techniques can be used to defame persons by replacing their face by the face of a different person". That is a definition by harm, and it is the one the dictionaries inherited: all four entries define the word by misrepresentation of a person, not by the architecture used. The paper also splits facial manipulation into "facial expression manipulation and facial identity manipulation", which is the seed of the taxonomy that later work expanded.

04How did the word widen to four kinds?

By 2024 the research vocabulary had settled on four kinds. The DF40 paper states that "deepfake techniques can be typically classified into four types: face-swapping (FS), face-reenactment (FR), entire face synthesis (EFS), and face editing (FE)", and builds its dataset from 10, 13, 12 and 5 methods of each respectively (Yan et al., arXiv 2406.13495). Its first footnote gives the authors' own gloss on the word: "Deepfake is a term derived from 'deep' (the use of deep nets) and 'fake' (does not exist)".

The widening matters for detection because a detector trained on one kind does not transfer to the others. The kinds of face deepfake page describes each kind; the DF40 dataset page carries the paper's tables, in which a model trained on face swaps scored 0.810 on entire face synthesis and one trained on synthesis scored 0.665 on swaps, both within the same data domain. When the word "deepfake" appears in a benchmark title, the first question is which of the four kinds the test set contains.

05Does deepfake include voices and still images?

In the dictionaries, yes. Merriam-Webster's definition opens with "a digital image, video, voice, etc.", Oxford's with "a video or audio recording", and Dictionary.com's with "video or audio file". Official usage agrees: the Hong Kong Government's working definition in a June 2024 Legislative Council reply is "the use of deep learning (namely a technique of artificial intelligence (AI)) to synthesise images for the purpose of creating falsified images", and the US FCC's ruling of 8 February 2024 treats cloned voices as "artificial" voices under telephone law. Two of the four incidents on the documented deepfake incidents page involve a cloned voice and no face at all.

In the face detection literature the word is narrower by habit. The datasets behind most published scores contain swaps and reenactments of real footage, and entirely generated faces were long filed under a separate heading, which is why DF40's inclusion of 12 synthesis methods was a change. Voice is a separate field with its own benchmarks. A reader should expect "deepfake" to mean face video in a detection paper and any synthetic likeness in a dictionary or a regulator's text, and should check which is meant before comparing numbers.

06How should the word be used on a benchmark page?

This site uses the broad sense in prose and the narrow sense only when quoting a paper's method name, written DeepFakes with the capital letter the FaceForensics++ authors chose. A page that reports a score names the kind of deepfake in the test set rather than relying on the word to carry it, because the word no longer does, and it names the metric beside the kind, since a score only has a meaning at a point on the ROC curve for deepfake detection. The FaceForensics++ authors needed a sentence to separate the two senses in 2019; by 2024 the four-kind taxonomy had replaced the word altogether as the unit of comparison.

One cited caveat belongs here. The dictionaries define a deepfake by its effect on a person, the papers define it by the technique, and neither definition says anything about how convincing a given example is: a crude swap by a trained model is a deepfake under both, and a convincing edit made by hand is a deepfake under neither.