Explainer
Deepfakes explained: how they're made, spotted and regulated
Deepfakes are synthetic videos, images, audio or text made with AI to show events that never happened.[1] Detection is a cat-and-mouse game, so provenance tools such as C2PA Content Credentials and watermarks are used alongside detectors.[2][3] In the US, the TAKE IT DOWN Act made publishing nonconsensual intimate deepfakes a federal crime and has required platforms to remove them within 48 hours since May 2026; the EU AI Act requires deepfakes to be disclosed.[4][5][6]
A deepfake is a video, picture, audio clip or text made with artificial intelligence to show something that never happened.[1] The same generative tools that power AI image and video generators can be used to make them, which is why laws in the US and EU now target their misuse.[7][4][6]
This explainer covers techniques, detection and the main laws in the US and EU as of November 2026. It is not legal advice, and it does not describe how to make a deepfake step by step.
How deepfakes are made
One common technique is the face swap: putting one person’s face onto another person’s body.[8] Another is lip-syncing, which matches a voice recording to a video so a person appears to say words they never said.[9] Voice cloning is also used in scams: in 2025 the FBI warned that criminals were sending AI-generated voice messages pretending to be senior US officials.[10]
A 2021 report from a Department of Homeland Security public-private exchange programme describes the main methods. Face-swap models built on autoencoders map images of person A and person B into the same compressed latent space with shared encoder parameters.[8] Generative adversarial networks (GANs) pair a generator that makes examples with an “adversary” network that tries to spot their flaws.[11] Newer diffusion models build images through a sequence of denoising steps, often in a compressed latent space conditioned on text.[7][12]
Not every fake is sophisticated. The same DHS report describes “cheapfakes”, in which simple digital techniques are applied to content to change how viewers perceive an event.[9] The DHS report argues that fakes work because people tend to believe what they see, so they need not be especially advanced to spread false information.[13]
Why detection is hard
NIST’s 2024 report on synthetic content sorts detection into three overlapping approaches: looking for provenance data such as watermarks or metadata, automated analysis of traces left by the generator, and human review.[14]
Detection tools are always catching up. NIST says that as soon as a new detector appears, generators improve, and a detector trained on one AI tool may miss fakes from another.[2] People are not a reliable backstop either: human checking is slow and costly, and gets harder as fakes improve.[15]
NIST describes automated content-based detection as a constant cat-and-mouse game between the detection and generation communities, and notes that detectors are often tied to specific generators.[2] Watermark-based detection has its own error rates: a covert watermark must be read by a detector with some nonzero probability of false positives and false negatives.[16] Human-in-the-loop review is labour-intensive and assumes people can detect synthetic content at all.[15]
Provenance and watermarks
Instead of judging content after the fact, provenance tools record where it came from. The Coalition for Content Provenance and Authenticity (C2PA) publishes an open standard called Content Credentials.[3] A Content Credential is a cryptographically bound record of an asset’s origin and edits, such as when it was created and how it was changed.[17] The C2PA itself notes that this metadata can be stripped and that adding it is optional; “durable” credentials add a watermark or fingerprint to recover it.[18]
Google DeepMind’s SynthID embeds imperceptible watermarks in AI-generated images, audio, text and video, and Google offers a detector portal to check content from its tools and partners.[19][20]
US laws
The TAKE IT DOWN Act (Public Law 119-12) was approved on May 19, 2025.[21] It makes it a federal crime to knowingly publish, without consent, a “digital forgery”: an intimate image of an identifiable person made with AI or other technology that a reasonable person would find indistinguishable from a real one.[4] Offences involving adults carry up to 2 years in prison, and those involving minors up to 3 years.[22]
The law also gave covered platforms one year to set up a notice-and-removal process, and requires them to remove a reported image, and known identical copies, within 48 hours of a valid request.[23] The Federal Trade Commission enforces these platform rules. A warning-letter template from its consumer protection bureau, dated May 20, 2026, the day after the deadline, says violators may face civil penalties of up to $53,088 per violation.[5]
States have their own laws. Minnesota, for example, makes it a crime to disseminate a deep fake to influence an election without the depicted person’s consent, and lets people sue over sexual deep fakes made without their consent.[24][25] Public Citizen, an advocacy group that tracks state election-deepfake bills, says such laws have passed with bipartisan support in every state that adopted one.[26] AI image tools have also drawn lawsuits: SpaceX reports that it has been named in suits alleging that xAI’s Grok enabled nonconsensual explicit images.[27]
EU rules
The EU AI Act requires people to be told when content is a deepfake.[28] Article 50(4) requires deployers of systems that generate or manipulate deepfake image, audio or video content to disclose that it is artificial, with lighter rules for evidently artistic, satirical or fictional works.[6] Article 50(2) requires providers of generative systems to mark output in a machine-readable format.[29] These rules took effect in August 2026, and breaches can bring fines of up to 15 million euros or 3% of worldwide turnover.[30][31] The EU’s Omnibus amendments also add a ban on AI systems that generate nonconsensual intimate content or child sexual abuse material, applying from 2 December 2026.[32] The EU AI Act explainer covers the wider law.
Read next
xAI
xAI, now part of SpaceX, makes the Grok AI models and runs the Colossus data centres in Memphis. Its models, deals and permits as of November 2026.
Questions readers ask
What is a deepfake?
A deepfake is synthetic media made with AI and machine learning that shows events that never happened, in video, images, audio or text. A related category, "cheapfakes", uses simple digital techniques to alter how viewers perceive an event.[1][9]
How are deepfakes made?
Common techniques include face swaps, which can use autoencoders that learn two faces in a shared compressed space; generative adversarial networks, where one network generates and another hunts for flaws; and lip-syncing, which maps a voice recording onto video.[8][11][9]
Can software reliably detect deepfakes?
Not reliably on its own. NIST calls automated detection a cat-and-mouse game and says detectors often work well only on the generators they were built for. Watermark detectors also have some rate of false positives and false negatives.[2][16]
Is it illegal to make or share a deepfake in the US?
It depends on the content. The federal TAKE IT DOWN Act makes it a crime to knowingly publish an intimate "digital forgery" of an identifiable person without consent, with up to 2 years in prison for offences involving adults and 3 years for minors. Some states, such as Minnesota, also restrict election deepfakes.[4][22][24]
How fast must platforms remove intimate deepfakes?
Under the TAKE IT DOWN Act, within 48 hours of a valid request. Platforms had to have a removal process by May 19, 2026, and the FTC enforces the rule, with civil penalties of up to $53,088 per violation.[23][5]
Does the EU require deepfakes to be labelled?
Yes. Article 50 of the AI Act requires deployers to disclose deepfake image, audio or video content, with lighter rules for evidently artistic or satirical works, and requires providers to mark synthetic output in a machine-readable way. The transparency rules took effect in August 2026.[6][29][30]
Sources
· 32 claims from 13 publishers
Each numbered claim is a statement we checked against the sources listed with it. Tap a number in the text to see its sources in place.
- [1]
A 2021 report from the US Department of Homeland Security's Public-Private Analytic Exchange Program describes deepfakes as a form of synthetic media that uses AI and machine learning to create believable videos, pictures, audio and text of events that never happened. reportedas of 2021-09-15
- Increasing Threat of Deepfake Identities · US Department of Homeland Security, Public-Private Analytic Exchange Program · Secondary source · 2021-09-15 · Abstract, p. 3 (retrieved 2026-11-07)
- [2]
NIST describes automated detection as a constant cat-and-mouse game in which generators improve as soon as new detectors appear, and says detectors often perform well only on the specific generators they were built for. confirmedas of 2024-11-20
- NIST AI 100-4: Reducing Risks Posed by Synthetic Content · National Institute of Standards and Technology · Primary source · 2024-11-20 · Section 4 (retrieved 2026-11-07)
- NIST AI 100-4: Reducing Risks Posed by Synthetic Content · National Institute of Standards and Technology · Primary source · 2024-11-20 · Section 4 (retrieved 2026-11-07)
- [3]
The Coalition for Content Provenance and Authenticity (C2PA) publishes an open technical standard, called Content Credentials, for establishing the origin and edits of digital content. confirmedas of 2026-11-07
- C2PA | Verifying Media Content Sources · Coalition for Content Provenance and Authenticity · Primary source (retrieved 2026-11-07)
- [4]
The TAKE IT DOWN Act makes it a federal crime to knowingly publish, without consent, a "digital forgery", an intimate image of an identifiable person created with AI or other technological means that a reasonable person would find indistinguishable from an authentic depiction. confirmedas of 2025-05-19
- Public Law 119-12, TAKE IT DOWN Act · US Government Publishing Office (govinfo) · Primary source · 2025-05-19 · Sec. 2, definitions (retrieved 2026-11-07)
- Public Law 119-12, TAKE IT DOWN Act · US Government Publishing Office (govinfo) · Primary source · 2025-05-19 · Sec. 2, definitions (retrieved 2026-11-07)
- Public Law 119-12, TAKE IT DOWN Act · US Government Publishing Office (govinfo) · Primary source · 2025-05-19 · Sec. 2, offense involving digital forgeries (retrieved 2026-11-07)
- [5]
The FTC enforces the platform provisions of the TAKE IT DOWN Act; a warning-letter template from its Bureau of Consumer Protection dated May 20, 2026 says covered platforms had been required to comply by May 19, 2026 and that violators may face civil penalties of up to $53,088 per violation. confirmedas of 2026-05-20
- FTC warning letter template regarding the TAKE IT DOWN Act · Federal Trade Commission · Primary source · 2026-05-20 · p. 1 (retrieved 2026-11-07)
- FTC warning letter template regarding the TAKE IT DOWN Act · Federal Trade Commission · Primary source · 2026-05-20 · Letter template dated May 20, 2026, p. 1 (retrieved 2026-11-07)
- FTC warning letter template regarding the TAKE IT DOWN Act · Federal Trade Commission · Primary source · 2026-05-20 · p. 2 (retrieved 2026-11-07)
- [6]
Article 50(4) requires deployers of AI systems that generate or manipulate deep-fake image, audio or video content to disclose that it is artificially generated or manipulated, with lighter rules for evidently artistic, satirical or fictional works. confirmedas of 2024-07-12
- Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) · EUR-Lex (Official Journal of the European Union) · Primary source · 2024-07-12 · Article 50(4) (retrieved 2026-11-07)
- Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) · EUR-Lex (Official Journal of the European Union) · Primary source · 2024-07-12 · Article 50(4) (retrieved 2026-11-07)
- [7]
Diffusion models build an image through a sequence of denoising steps, and working directly on pixels made training powerful models cost hundreds of GPU days. confirmedas of 2021-12-20
- High-Resolution Image Synthesis with Latent Diffusion Models (Rombach et al.) · arXiv · Primary source · 2021-12-20 · Abstract (retrieved 2026-11-07)
- High-Resolution Image Synthesis with Latent Diffusion Models (Rombach et al.) · arXiv · Primary source · 2021-12-20 · Abstract (retrieved 2026-11-07)
- [8]
The DHS exchange-program report says a face swap places one person's face onto another person's body, and that an autoencoder approach maps images of both people into the same compressed latent space using shared encoder parameters. reportedas of 2021-09-15
- Increasing Threat of Deepfake Identities · US Department of Homeland Security, Public-Private Analytic Exchange Program · Secondary source · 2021-09-15 · Face swap section (retrieved 2026-11-07)
- Increasing Threat of Deepfake Identities · US Department of Homeland Security, Public-Private Analytic Exchange Program · Secondary source · 2021-09-15 · Face swap section (retrieved 2026-11-07)
- [9]
The DHS exchange-program report also describes lip-sync deepfakes, which map a voice recording onto video so the subject appears to say something, and "cheapfakes", in which simple digital techniques are applied to content to alter how viewers perceive an event. reportedas of 2021-09-15
- Increasing Threat of Deepfake Identities · US Department of Homeland Security, Public-Private Analytic Exchange Program · Secondary source · 2021-09-15 · Lip syncing section (retrieved 2026-11-07)
- Increasing Threat of Deepfake Identities · US Department of Homeland Security, Public-Private Analytic Exchange Program · Secondary source · 2021-09-15 · Lip syncing section (retrieved 2026-11-07)
- Increasing Threat of Deepfake Identities · US Department of Homeland Security, Public-Private Analytic Exchange Program · Secondary source · 2021-09-15 (retrieved 2026-11-07)
- [10]
In May 2025 the FBI warned that since April 2025 malicious actors had impersonated senior US officials using text messages and AI-generated voice messages. confirmedas of 2025-05-15
- Senior US Officials Impersonated in Malicious Messaging Campaign (Alert I-051525-PSA) · FBI Internet Crime Complaint Center · Primary source · 2025-05-15 (retrieved 2026-11-07)
- Senior US Officials Impersonated in Malicious Messaging Campaign (Alert I-051525-PSA) · FBI Internet Crime Complaint Center · Primary source · 2025-05-15 (retrieved 2026-11-07)
- [11]
The DHS exchange-program report describes generative adversarial networks (GANs), in which a generator network creates examples and a second "adversary" network tries to detect their flaws, as a key technology for producing deepfakes. reportedas of 2021-09-15
- Increasing Threat of Deepfake Identities · US Department of Homeland Security, Public-Private Analytic Exchange Program · Secondary source · 2021-09-15 · Generative Adversarial Networks box (retrieved 2026-11-07)
- Increasing Threat of Deepfake Identities · US Department of Homeland Security, Public-Private Analytic Exchange Program · Secondary source · 2021-09-15 · Generative Adversarial Networks box (retrieved 2026-11-07)
- [12]
Latent diffusion runs the diffusion process in the compressed latent space of a pretrained autoencoder and uses cross-attention layers to condition generation on inputs such as text. confirmedas of 2021-12-20
- High-Resolution Image Synthesis with Latent Diffusion Models (Rombach et al.) · arXiv · Primary source · 2021-12-20 · Abstract (retrieved 2026-11-07)
- High-Resolution Image Synthesis with Latent Diffusion Models (Rombach et al.) · arXiv · Primary source · 2021-12-20 · Abstract (retrieved 2026-11-07)
- [13]
The DHS exchange-program report argues that the threat comes from people's inclination to believe what they see, so deepfakes need not be especially advanced to spread misinformation. reportedas of 2021-09-15· interpretation
- Increasing Threat of Deepfake Identities · US Department of Homeland Security, Public-Private Analytic Exchange Program · Secondary source · 2021-09-15 · Abstract, p. 3 (retrieved 2026-11-07)
- [14]
NIST's November 2024 report on synthetic content groups detection techniques into three overlapping categories; detecting provenance data such as watermarks or metadata, automated content-based detection of traces left during generation, and human-assisted detection. confirmedas of 2024-11-20
- NIST AI 100-4: Reducing Risks Posed by Synthetic Content · National Institute of Standards and Technology · Primary source · 2024-11-20 · Section 4 (synthetic content detection) (retrieved 2026-11-07)
- NIST AI 100-4: Reducing Risks Posed by Synthetic Content · National Institute of Standards and Technology · Primary source · 2024-11-20 · Section 4 (retrieved 2026-11-07)
- [15]
NIST says human-assisted detection is labour-intensive and costly, and that people may find detection harder as synthetic content grows more sophisticated. confirmedas of 2024-11-20
- NIST AI 100-4: Reducing Risks Posed by Synthetic Content · National Institute of Standards and Technology · Primary source · 2024-11-20 · Section 4 (retrieved 2026-11-07)
- NIST AI 100-4: Reducing Risks Posed by Synthetic Content · National Institute of Standards and Technology · Primary source · 2024-11-20 · Section 4 (retrieved 2026-11-07)
- [16]
NIST notes that a covert watermark must be read by a detector that has some nonzero rate of false positives and false negatives. confirmedas of 2024-11-20
- NIST AI 100-4: Reducing Risks Posed by Synthetic Content · National Institute of Standards and Technology · Primary source · 2024-11-20 · Section 3 (digital watermarking) (retrieved 2026-11-07)
- [17]
A C2PA Content Credential is a cryptographically bound record of an asset's provenance containing assertions such as when and where it was created and how it was modified. confirmedas of 2026-11-07
- C2PA and Content Credentials Explainer (version 2.2) · Coalition for Content Provenance and Authenticity · Primary source · 1. Introduction (retrieved 2026-11-07)
- C2PA and Content Credentials Explainer (version 2.2) · Coalition for Content Provenance and Authenticity · Primary source · 1. Introduction (retrieved 2026-11-07)
- [18]
The C2PA explainer says provenance metadata can be removed, that "durable" credentials add watermarking or fingerprinting to recover it, and that adding provenance is optional. confirmedas of 2026-11-07
- C2PA and Content Credentials Explainer (version 2.2) · Coalition for Content Provenance and Authenticity · Primary source · 7.2.3 (retrieved 2026-11-07)
- C2PA and Content Credentials Explainer (version 2.2) · Coalition for Content Provenance and Authenticity · Primary source · 7.2.3 (retrieved 2026-11-07)
- C2PA and Content Credentials Explainer (version 2.2) · Coalition for Content Provenance and Authenticity · Primary source · 7.1.2 (retrieved 2026-11-07)
- [19]
Google DeepMind's SynthID embeds imperceptible watermarks in AI-generated images, audio, text and video, and Google says image and video watermarks are designed to survive cropping, filters and lossy compression. confirmedas of 2026-11-07
- [20]
Google launched the SynthID Detector verification portal in 2025 and says it is expanding access to everyone, to check content made with AI from Google or partners including OpenAI, NVIDIA and Kakao. confirmedas of 2026-11-07
- [21]
The TAKE IT DOWN Act (Public Law 119-12) was approved on May 19, 2025. confirmedas of 2025-05-19
- Public Law 119-12, TAKE IT DOWN Act · US Government Publishing Office (govinfo) · Primary source · 2025-05-19 (retrieved 2026-11-07)
- [22]
Under the TAKE IT DOWN Act, publication offences involving adults carry up to 2 years in prison and those involving minors up to 3 years. confirmedas of 2025-05-19
- Public Law 119-12, TAKE IT DOWN Act · US Government Publishing Office (govinfo) · Primary source · 2025-05-19 · Sec. 2, penalties (retrieved 2026-11-07)
- Public Law 119-12, TAKE IT DOWN Act · US Government Publishing Office (govinfo) · Primary source · 2025-05-19 · Sec. 2, penalties (retrieved 2026-11-07)
- [23]
The TAKE IT DOWN Act requires covered platforms to set up a notice-and-removal process within one year of enactment and to remove a reported nonconsensual intimate image, and make reasonable efforts to remove known identical copies, within 48 hours of a valid request. confirmedas of 2025-05-19
- Public Law 119-12, TAKE IT DOWN Act · US Government Publishing Office (govinfo) · Primary source · 2025-05-19 · Sec. 3(a)(1)(A) (retrieved 2026-11-07)
- Public Law 119-12, TAKE IT DOWN Act · US Government Publishing Office (govinfo) · Primary source · 2025-05-19 · Sec. 3(a)(3) (retrieved 2026-11-07)
- [24]
Minnesota law (section 609.771, enacted in 2023 and amended in 2024) makes it a crime to disseminate a deep fake to influence an election without the depicted person's consent, under conditions set out in the statute. confirmedas of 2026-11-07
- Minnesota Statutes 609.771, Use of deep fake technology to influence election · Minnesota Office of the Revisor of Statutes · Primary source (retrieved 2026-11-07)
- Minnesota Statutes 609.771, Use of deep fake technology to influence election · Minnesota Office of the Revisor of Statutes · Primary source · Subd. 2 (retrieved 2026-11-07)
- [25]
Minnesota law (section 604.32) gives people a civil cause of action against someone who disseminates a sexual deep fake of them knowing they did not consent. confirmedas of 2026-11-07
- Minnesota Statutes 604.32, Nonconsensual dissemination of deep fake depicting intimate parts or sexual acts · Minnesota Office of the Revisor of Statutes · Primary source · Subd. 2 (retrieved 2026-11-07)
- [26]
Public Citizen, an advocacy group, tracks state bills on deepfakes in elections and, in a tracker updated October 21, 2026, said such laws had passed with bipartisan support in every state that enacted one. reportedas of 2026-10-21
- Tracker: State Legislation on Deepfakes in Elections · Public Citizen · Secondary source · 2026-10-21 (retrieved 2026-11-07)
- Tracker: State Legislation on Deepfakes in Elections · Public Citizen · Secondary source · 2026-10-21 (retrieved 2026-11-07)
- [27]
SpaceX reports that since January 2026 it and some subsidiaries have been named in lawsuits alleging that Grok's image-generation and editing features enabled nonconsensual explicit images. confirmedas of 2026-10-22
- Space Exploration Technologies Corp. Form 10-Q for the quarter ended September 30, 2026 · SpaceX (SEC filing) · Primary source · 2026-10-22 · Legal proceedings (retrieved 2026-11-07)
- Space Exploration Technologies Corp. Form 10-Q for the quarter ended September 30, 2026 · SpaceX (SEC filing) · Primary source · 2026-10-22 · Legal proceedings (retrieved 2026-11-07)
- [28]
Under the AI Act's transparency rules, people must be made aware when they are interacting with an AI system such as a chatbot, and certain AI-generated content, namely deep fakes and text published to inform the public on matters of public interest, must be clearly labelled. confirmedas of 2026-11-07
- AI Act (regulatory framework for AI) · European Commission · Primary source · Transparency risk (retrieved 2026-11-07)
- AI Act (regulatory framework for AI) · European Commission · Primary source · Transparency risk (retrieved 2026-11-07)
- [29]
Article 50(2) requires providers of AI systems that generate synthetic audio, image, video or text to mark the outputs in a machine-readable format so they are detectable as artificially generated or manipulated. confirmedas of 2024-07-12
- Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) · EUR-Lex (Official Journal of the European Union) · Primary source · 2024-07-12 · Article 50(2) (retrieved 2026-11-07)
- [30]
The AI Act's transparency rules took effect in August 2026, when the Act became generally applicable on 2 August 2026. confirmedas of 2026-08-02
- AI Act (regulatory framework for AI) · European Commission · Primary source · Transparency risk (retrieved 2026-11-07)
- AI Act (regulatory framework for AI) · European Commission · Primary source · Application timeline (retrieved 2026-11-07)
- [31]
Breaches of most other operator obligations, including the Article 50 transparency rules, can bring fines of up to 15 million euros or 3% of worldwide turnover, and supplying incorrect or misleading information to authorities up to 7.5 million euros or 1%. confirmedas of 2024-07-12
- Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) · EUR-Lex (Official Journal of the European Union) · Primary source · 2024-07-12 · Article 99(4) (retrieved 2026-11-07)
- Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) · EUR-Lex (Official Journal of the European Union) · Primary source · 2024-07-12 · Article 99(4)(g) (retrieved 2026-11-07)
- Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) · EUR-Lex (Official Journal of the European Union) · Primary source · 2024-07-12 · Article 99(5) (retrieved 2026-11-07)
- [32]
The Omnibus added a prohibition on AI systems that generate non-consensual sexually explicit and intimate content or child sexual abuse material, which applies from 2 December 2026. confirmedas of 2026-07-27
- AI Omnibus enters into force · European Commission · Primary source · 2026-07-27 · Key changes (retrieved 2026-11-07)
- Regulation (EU) 2026/1744 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 (Digital Omnibus on AI) · EUR-Lex (Official Journal of the European Union) · Primary source · 2026-07-24 · Article 1(40), amending Article 113(a) (retrieved 2026-11-07)
Revision history (1)
- Page created.
Created Nov 7, 2026. Last reviewed by an editor on Nov 8, 2026. Next scheduled review: Mar 7, 2027.
Cite this page
"Deepfakes explained: how they're made, spotted and regulated." ContentLora, updated Nov 7, 2026. https://contentlora.com/explain/deepfakes-explained
Spotted an error? Suggest a correction or emailcorrections@contentlora.com.