deepfake
A video, image or audio clip, made or altered with AI, that convincingly shows a real person saying or doing something they never said or did.
- A cloned presidential voice telling voters to skip a primary
- A video call in which the finance chief and colleagues are fakes
- A video of a wartime leader telling his troops to surrender
- Intimate images of a real person made without consent
- Earliest AI use
- 2017
- A Reddit user posting as deepfakes claims a face-swapped video.
- Current meaning
- 2018
- Widened to any AI-faked audio or video of real people.
- Coined by
- A Reddit user
- Anonymous. The username joins deep learning and fake.
- Also called
- Deep fake
- The two-word form used in EU law and early legal scholarship.
In plain words
In AI, a deepfake is fabricated video, audio or imagery that shows a real person saying or doing something they never did, made to pass as genuine.1
Deepfakes are made with deep learning, the branch of machine learning behind generative AI. A system trained on images of a person can swap that face into other footage or change its expression. Trained on recordings, it can copy the person’s voice.2
The first deepfakes, in late 2017, were sexual videos of celebrities made without their consent. In 2019 the company Deeptrace counted 14,678 deepfake videos online, and 96% were non-consensual pornography. Politics and fraud followed. In March 2022 a faked video showed Ukraine’s president, Volodymyr Zelenskyy, telling his soldiers to surrender. In January 2024 a robocall imitating President Biden told New Hampshire voters to stay home. That year Arup, a British design firm, also confirmed a deepfake fraud. An employee in Hong Kong had sent $25.6 million to criminals after a video call with a faked finance chief.3
Law professors Robert Chesney and Danielle Citron describe a second harm, the liar’s dividend. Once people know that recordings can be faked, a real recording becomes easier to deny.4
A deepfake is not the same as a cheapfake, also called a shallow fake. A cheapfake misleads with ordinary tools, such as photo editing, a slowed-down clip or a lookalike.5 A deepfake also differs from AI slop, cheap content made in volume that need not show anyone real.
How it happens
The systems behind deepfakes are neural networks trained on examples of a person’s face or voice. In a face swap, one person’s face replaces another’s in existing video. Networks called autoencoders learn a person’s facial features and rebuild the face with new expressions. Generative adversarial networks, or GANs, learn to produce new images of the same quality as their training images. By 2021, a few minutes of recorded speech could be enough to clone a voice.2
Spotting the result is hard. Detection software looks for visual flaws and for inconsistencies between video frames. It struggles with fakes made by methods it has not seen. In a 2020 contest run by Facebook, none of 2,114 entrants reached 70% accuracy on unseen deepfakes. In a 2023 study, listeners caught speech deepfakes 73% of the time. Platforms often compress shared files, which interferes with detection, and even simple edits can make a detector unreliable.2
A second approach tracks the origin of genuine files. The C2PA is a coalition that includes Adobe, Google, Microsoft and OpenAI. Its open standard records where a file came from and how it was edited.6 Laws also require labels. From 2 August 2026, the EU AI Act requires anyone who uses AI to make a deep fake to disclose that it is artificial. Personal, non-professional use is exempt.
Origin and first known use
A blend of deep, from deep learning, and fake. The word began as a username. In late 2017 an anonymous Reddit user calling himself deepfakes claimed to have made a face-swapped celebrity video with deep learning. The first such videos were sexual and made without consent. By mid-2018 law professors were using deep fake for any faked audio or video of real people. EU law keeps that two-word spelling. The US TAKE IT DOWN Act calls intimate AI fakes digital forgeries.
- 2017 An anonymous Reddit user calling himself deepfakes claims to have made a face-swapped celebrity video. The name comes from the deep learning behind it.US Department of Homeland Security, “Increasing Threats of Deepfake Identities” (undated) dhs.gov →: source for 2017
- 2018 Law professors Robert Chesney and Danielle Citron use deep fake for any faked audio or video of real people, and warn of risks to democracy.Chesney and Citron, “Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security”, SSRN, July 2018; California Law Review 107 (2019) ssrn.com →: source for 2018
- 2019 Deeptrace counts 14,678 deepfake videos online, almost double the number seven months earlier. Non-consensual pornography makes up 96% of them.Ajder, Patrini, Cavalli and Cullen, “The State of Deepfakes: Landscape, Threats, and Impact”, Deeptrace, September 2019 regmedia.co.uk →: source for 2019
- 2022 A deepfake of Ukraine’s president, Volodymyr Zelenskyy, telling soldiers to surrender appears on a hacked Ukrainian news site. Facebook, YouTube and Twitter remove it.NPR report on the Zelenskyy deepfake, 16 March 2022 npr.org →: source for 2022
- 2024 After a robocall imitating President Biden told New Hampshire voters to stay home, the FCC rules that AI-generated voices are “artificial” under US robocall law.Federal Communications Commission, Declaratory Ruling FCC 24-17, adopted 2 February 2024, released 8 February 2024 fcc.gov →: source for 2024
- 2024 The EU AI Act requires those who deploy AI systems that make deep fakes to disclose that the content is artificially generated or manipulated.Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 50(4), 13 June 2024 ai-act-service-desk.ec.europa.eu →: source for 2024
- 2025 The US TAKE IT DOWN Act criminalizes knowingly publishing non-consensual intimate images, including AI-made digital forgeries, and requires removal within 48 hours of a request.TAKE IT DOWN Act, Public Law 119-12, 19 May 2025 govinfo.gov →: source for 2025
- 2026 EU duties to mark and label AI-generated content, including deep fakes, apply from 2 August 2026, backed by a code of practice finalized in June.European Commission, “Code of Practice on Transparency of AI-generated Content”, page last updated 31 July 2026 digital-strategy.ec.europa.eu →: source for 2026
In the wild
-
“…the ability to distort reality has taken an exponential leap forward with ‘deep fake’ technology.”
Chesney and Citron, SSRN preprint (California Law Review, 2019) ssrn.com →: Chesney and Citron, SSRN preprint (California Law Review, 2019), 21 Jul 2018Archived →: copy of Chesney and Citron, SSRN preprint (California Law Review, 2019), 21 Jul 2018 -
“A ‘deepfake’ is a video that has been altered through some form of machine learning to ‘hybridize or generate human bodies and faces,’…”
Paris and Donovan, Data & Society report datasociety.net →: Paris and Donovan, Data & Society report, 18 Sep 2019Archived →: copy of Paris and Donovan, Data & Society report, 18 Sep 2019 -
“None of the 2,114 participants… achieved 70 percent accuracy on unseen deepfakes in the black box dataset.”
Facebook AI, Deepfake Detection Challenge results ai.meta.com →: Facebook AI, Deepfake Detection Challenge results, 12 Jun 2020Archived →: copy of Facebook AI, Deepfake Detection Challenge results, 12 Jun 2020 -
“Deepfake video of Zelenskyy could be ‘tip of the iceberg’ in info war, experts warn”
-
“Listeners only correctly spotted the deepfakes 73% of the time…”
-
“A deepfake ‘CFO’ tricked the British design firm behind the Sydney Opera House in $25 million scam”
-
‘deep fake’ means AI-generated or manipulated image, audio or video content that resembles existing persons… and would falsely appear to a person to be authentic…”
-
“…cited as the “Tools to Address Known Exploitation by Immobilizing Technological Deepfakes on Websites and Networks Act” or the “TAKE IT DOWN Act”.”
US Congress, TAKE IT DOWN Act (Public Law 119-12) govinfo.gov →: US Congress, TAKE IT DOWN Act (Public Law 119-12), 19 May 2025Archived →: copy of US Congress, TAKE IT DOWN Act (Public Law 119-12), 19 May 2025 -
“Certain AI-generated content should be clearly and visibly labelled, namely deep fakes…”
European Commission, AI Act policy page digital-strategy.ec.europa.eu →: European Commission, AI Act policy page, 3 Aug 2026Archived →: copy of European Commission, AI Act policy page, 3 Aug 2026
Not to be confused with
-
Cheapfake
A cheapfake, or shallow fake, misleads with ordinary tools such as photo editing, a slowed-down clip or a lookalike. A deepfake uses deep learning to generate or alter a face or voice.
-
AI slop
AI slop is cheap, low-effort content made in volume to fill feeds. A deepfake is made to deceive about a real person or event, and slop need not show anyone real.
-
Synthetic media
Synthetic media covers all AI-made or AI-altered content, including dubbing, avatars and effects made with consent and shown as synthetic. A deepfake is synthetic media passed off as a genuine recording of a real person.
Questions
What is a deepfake?
A deepfake is a video, image or audio clip made or altered with AI. It makes a real person appear to say or do something they never said or did. The word joins deep learning and fake. It began in late 2017 as the name of an anonymous Reddit user. Deepfakes now appear in fraud, political disinformation and sexual images made without consent.
How can a deepfake be detected?
Reliable detection is hard. In a 2020 contest run by Facebook, none of 2,114 entrants built a detector that reached 70% accuracy on deepfakes it had not seen before. In a 2023 study, listeners caught speech deepfakes 73% of the time. Companies including Adobe, Google, Microsoft and OpenAI also back the C2PA standard, which records where a file came from and how it was edited.
Are deepfakes illegal?
It depends on the use and the country. The US TAKE IT DOWN Act of May 2025 makes it a federal crime to knowingly publish intimate images without consent. That includes AI-made ones. Platforms must remove them within 48 hours of a valid request. The FCC ruled in 2024 that AI-generated voices in robocalls count as artificial. The EU AI Act requires deep fakes to be labeled from 2 August 2026.
What is the difference between a deepfake and a cheapfake?
A deepfake uses deep learning to generate or alter a person’s face or voice. A cheapfake, also called a shallow fake, uses older and cheaper methods, such as photo editing, slowing down or speeding up video, or a lookalike. It needs less skill to make. Britt Paris and Joan Donovan of Data & Society set out the contrast between the two in a 2019 report.
Where does the word deepfake come from?
It joins deep, from deep learning, with fake. In late 2017 an anonymous Reddit user calling himself deepfakes claimed to have made a face-swapped celebrity video with deep learning. The username became the name of the technique. By mid-2018 law professors Robert Chesney and Danielle Citron used it for any faked audio or video of real people. EU law spells it deep fake.
Watch
- How deepfakes undermine truth and threaten democracyLaw professor Danielle Citron on how deepfakes harm people and public trust, and what can be done.
- When AI can fake reality, who can you trust?Human rights technologist Sam Gregory sets out three steps to help people tell real media from synthetic.
Cite this entry
“Deepfake.” SUPERR_dictionary, Superrduperr, 24 Sept. 2026, superrduperr.com/
Superrduperr. (2026, September 24). Deepfake. In SUPERR_dictionary. https://superrduperr.com/
Superrduperr. “Deepfake.” SUPERR_dictionary. September 24, 2026. https://superrduperr.com/