What is a deepfake?
Definition and typology, the model families behind fakes, what detection can and cannot do, and the US statutes that apply in 2026.
What counts as a deepfake?
A deepfake is a photo, video or audio clip that has been generated or altered by machine learning to show a real, identifiable person doing or saying something they did not do. The word appeared in 2017 as a Reddit username; by 2024 a US House hearing cited that 98% of deepfake videos online were pornographic, and 99% targeted women.
| Type | How it is made | Typical tells | US legal hook |
|---|---|---|---|
| Face swap | Encoder–decoder or one-shot swapper places a face onto a target body | Skin-tone seams, flat lighting, blinking errors | TAKE IT DOWN Act if intimate; state NCII laws |
| Body replacement (undress AI) | Diffusion inpainting fills a masked clothing region | Anatomy mismatches, blurred edges | TAKE IT DOWN Act “digital forgery”; 15 U.S.C. §6851 |
| Lip-sync / puppeteering | Audio-driven models move the mouth or head | Teeth artifacts, unnatural head stillness | State election and fraud statutes |
| Full synthesis | Text-to-image or text-to-video model with a reference image | No source photo; provenance data absent | Same as above when the subject is identifiable |
How are deepfakes made?
Two model families do most of the work. GANs, introduced in 2014, trained a generator against a discriminator and powered the first face-swap programs such as DeepFaceLab, archived in November 2024.
Diffusion models, released as Stable Diffusion in August 2022, now dominate image work because they follow prompts and masks better. Every method invents the missing information; none reveals anything true about the subject.
Deepfakes cause unique harms that are disproportionately experienced by women.
— Dr. Waldman, witness testimony at the US House hearing “Addressing Real Harm Done by Deepfakes”, March 12, 2024, accessed September 3, 2026
How reliable is deepfake detection?
Detectors are classifiers trained on datasets such as FaceForensics++ (2019, 1.8 million manipulated frames). They degrade on new models, compression and re-encoding.
Provenance systems such as C2PA Content Credentials take the opposite route: a signed manifest records how a file was made, but the signal disappears in a screenshot. Our deepfake detection entry lists tools and reported error rates.
Which US laws apply to sexual deepfakes in 2026?
The TAKE IT DOWN Act, Public Law 119-12, signed May 19, 2025, criminalizes knowing publication of a non-consensual intimate “digital forgery” of an identifiable person, with up to 2 years in prison (3 years for a minor), and requires covered platforms to remove reported material within 48 hours from May 19, 2026.
A civil action under 15 U.S.C. §6851 provides $150,000 in liquidated damages. The DEFIANCE Act (S.1837) passed the Senate and was pending in the House as of September 3, 2026; state statutes are collected in state deepfake laws.
What should a victim do first?
Save the URL and a dated screenshot, then file a removal request with the hosting platform under its TAKE IT DOWN procedure and hash the image with StopNCII.org (adults) or NCMEC’s Take It Down (under 18).
Anything hosted on this site goes through Report NCII; the full sequence is in Deepfake nudes: what to do.
Consent Note
This site is a publisher: it reviews third-party tools and does not create, edit, accept or store images. Docs entries are general information, not legal advice.
Sources
- TAKE IT DOWN Act, Public Law 119-12 — GovInfo, accessed September 3, 2026.
- Complying With the Take It Down Act — Federal Trade Commission, accessed September 3, 2026.
- 15 U.S.C. §6851 — Cornell LII, accessed September 3, 2026.
- 18 U.S.C. §2256 — Cornell LII, accessed September 3, 2026.
- S.1837 – DEFIANCE Act of 2025 (status) — Congress.gov, accessed September 3, 2026.
- Hearing “Addressing Real Harm Done by Deepfakes” (March 12, 2024) — U.S. House via GovInfo, accessed September 3, 2026.
FAQ
Is every AI-edited image a deepfake?
No. The term applies when a real, identifiable person is shown doing or saying something they did not do. A fully fictional face or a stylized self-portrait is generated media but not a deepfake. EU AI Act Article 50 labeling duties apply to any synthetic image.
Are deepfakes illegal in the United States?
Not as such. Parody, film effects and consented edits are lawful. Non-consensual intimate deepfakes of identifiable people are a federal offense under the TAKE IT DOWN Act since May 19, 2025, and sexual images of minors are prosecuted under 18 U.S.C. §2256 regardless of how they were produced.
How can I tell if a video is a deepfake?
Look for skin-tone seams around the face, teeth and earring artifacts, mismatched lighting between face and neck and unnatural stillness of the head. Check for C2PA Content Credentials or a SynthID mark. None of these signals is conclusive, and detectors miss newer methods.
What is a “digital forgery” under federal law?
The TAKE IT DOWN Act defines it as an intimate visual depiction of an identifiable individual created or altered by artificial intelligence or other technological means so that it is indistinguishable from an authentic depiction. Undress-app output and face-swap videos qualify when the subject is real.
Do reviewed undress tools produce deepfakes?
Output from these tools is a digital forgery whenever the subject is a real person who did not consent. Every tool in our rankings requires the user to confirm ownership or consent and blocks minors; tools without those controls are delisted.
How long do platforms have to remove a deepfake?
Since May 19, 2026, covered platforms must remove a reported non-consensual intimate image, including digital forgeries, within 48 hours of a valid request and make reasonable efforts to remove copies. The Federal Trade Commission enforces the duty with civil penalties of $53,088 per violation.