What is a deepfake?
What is a deepfake? It is synthetic media in which AI makes a real person appear to say or do something they did not. The deepfake meaning has widened over time: the word first described swapped faces in video, and it now covers cloned voices and fake photos too.
The word combines “deep learning”, the branch of AI behind the technique, and “fake”. In everyday use it means content that is both realistic and deceptive, not any image touched by AI.
The EU AI Act gives one of the clearest legal definitions. Article 3(60) of Regulation (EU) 2024/1689 defines a deep fake as AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful.
Two parts of that definition matter in daily life. The content resembles something real, and it would mislead someone into thinking it is authentic. A cartoon of a politician is not a deepfake; a realistic video of them giving a speech they never gave is.
What are the main types of deepfakes?
There are six main types of deepfakes, and they differ in what is faked: the face, the mouth, the movement, the whole scene, the voice or a still image. Knowing the type tells you where to look for flaws.
| Type | What is faked | Typical use | Where flaws show |
|---|---|---|---|
| Face swap | One person’s face placed onto another person’s body in a video | Impersonation, non-consensual content, parody | Face edges, skin tone at the jaw and hairline, glasses, profile views |
| Lip-sync | The mouth is altered to match new audio | Fake statements, dubbing | Teeth, inner mouth, sounds like “p”, “b” and “m” that need closed lips |
| Face reenactment | A real face is driven by another person’s expressions and head movements | Puppeted video of a public figure | Blinking, expression timing, head and body moving out of step |
| Fully generated video | The whole scene is created from a prompt or an image | Fake news footage, ads, entertainment | Physics, hands, text, objects that appear or vanish |
| Voice clone | A synthetic voice that sounds like a specific person | Phone scams, fake audio statements | Flat emotion, odd breathing, unusual pacing |
| Image deepfake | A still photo of a real person in a scene that never happened | Fake profile pictures, harassment, misinformation | Hands, text, lighting, reflections, backgrounds |
Not every manipulated clip is a deepfake. Slowing down a real video, cutting it out of context or giving it a false caption needs no AI at all. These cheap edits can mislead just as well, so check the context even when the pixels are genuine.
What is a deep fake image?
A deep fake image is a still picture that shows a real, identifiable person in a situation that never happened. It can be made by swapping a face into an existing photo, by editing part of a real picture, or by generating the whole scene from a prompt. Fake profile pictures are a common case; see how to spot a fake profile picture for the checks that work.
How are deepfakes made?
Deepfakes are made with machine learning models trained on many examples of faces, voices or video, which learn to produce new content in the same style. You do not need the engineering details to spot fakes, but a rough picture of the process explains where they tend to go wrong.
- Classic face swaps use models that learn to compress a face into a compact description and rebuild it with another person’s features, frame by frame.
- Generative adversarial networks, or GANs, pit a generator against a critic until the output looks realistic. They made convincing synthetic faces common.
- Diffusion models start from random noise and refine it step by step, guided by a text prompt or a reference image. Many recent image and video generators work this way.
- Voice cloning models learn the sound of a voice from recordings and then read any new text in that voice.
Each method leaves different traces. Frame-by-frame face swaps can flicker or drift at the edges of the face. Generated video can break the rules of physics or lose track of objects. Voice clones can sound fluent but emotionally flat. Generators improve quickly and these traces fade, which is why serious detection combines several methods.
The raw material is usually public. The more photos, videos and voice recordings of a person exist online, the easier it is to imitate them, which is why public figures are frequent targets. It is also why the FBI’s advice includes limiting how much of your image and voice you share publicly.
What are some common deepfake examples?
The most common deepfake examples follow a handful of patterns that repeat across countries and platforms. Recognising the pattern often matters more than inspecting the pixels, because the story around a fake is usually what gives it away.
- Celebrity endorsement scams: a famous face appears to recommend an investment, a product or a giveaway.
- Family emergency calls: a cloned voice of a relative asks for money urgently and asks you to keep it secret.
- Executive impersonation: a video call or voice message from a “boss” asks staff to pay an invoice or change bank details.
- Non-consensual intimate images: a person’s face is placed on explicit material without their consent.
- Political misinformation: a candidate appears to say something inflammatory shortly before a vote.
- Fake profiles: a generated face fronts a dating, recruiting or marketplace account.
The FBI’s Internet Crime Complaint Center described this trend in a public service announcement on December 3, 2024, titled “Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud”. It warns that criminals use generated images, video and cloned voices to make fraud more convincing. Our page on voice clone scams covers the phone version in detail.
Notice what these examples have in common. Each one asks you to do something quickly, whether that is paying, sharing, trusting a stranger or believing a scandal. The urgency is part of the technique, because it leaves no time for a second check.
Are all deepfakes harmful?
No. The same technology is used in film, accessibility and satire. What makes a deepfake harmful is deception and the lack of consent, not the technique itself.
| Legitimate uses | Harmful uses |
|---|---|
| Dubbing a film so actors’ lips match another language | Impersonating someone to steal money or data |
| Recreating a voice, with consent, for a person who lost theirs | Creating sexual content of a person without consent |
| Satire that is clearly labelled as such | Fabricating statements to sway an election |
| Training videos with a consenting presenter’s digital double | Harassing or blackmailing a private person |
| Visual effects in film and games | Faking evidence in a dispute, an insurance claim or a trial |
Rules are catching up. Article 50 of the EU AI Act, which applies from August 2, 2026, requires deployers of AI systems that create deepfakes to disclose that the content was artificially generated or manipulated, with a lighter duty for evidently artistic, creative, satirical or fictional works. In the United States, the TAKE IT DOWN Act, signed on May 19, 2025, targets non-consensual intimate images, including AI-generated ones, and the FTC began enforcing its platform removal duties on May 19, 2026.
How can you spot a deepfake?
You spot a deepfake by combining what you see and hear with where the content came from. Visual and audio clues help, but context checks catch many fakes that look and sound perfect.
Look
- Edges of the face, the hairline and the jaw, especially when the head turns.
- The mouth and teeth during speech, and whether lip shapes match the sounds.
- Blinking, gaze and expressions that feel delayed, frozen or repetitive.
- Hands, jewelry, glasses and text in the background.
- Light and shadows on the face compared with the rest of the scene.
Listen
- A voice that is fluent but flat, with little emotion or breathing.
- Odd pacing, clipped words or background noise that cuts in and out.
- Audio that does not match the room, the distance or the movement of the mouth.
Check the context
- Who posted it first, and do trusted outlets report the same thing?
- Does it push you to act fast: pay, share, sign or stay silent?
- Can you confirm through a channel you already trust, such as calling the person back on a number you know?
The FBI’s December 2024 announcement recommends agreeing on a secret word or phrase with family members to confirm identity, and hanging up and calling back on a number you look up yourself. For deeper checklists, see how to spot a deepfake video and how to tell if a voice is AI.
How does deepfake detection work?
Deepfake detection works by combining provenance checks with models trained to recognise the traces that generation leaves in pixels, motion and sound. No detector is perfect, so good tools explain their signals instead of giving a bare score.
- 01Provenance: C2PA Content Credentials or metadata may name the tool that made the file.
- 02Frame analysis: individual frames are scored for signs of generation.
- 03Face analysis: a classifier looks for the blending and texture traces typical of face swaps.
- 04Audio analysis: an anti-spoofing model scores the voice for signs of synthesis.
- 05Human review: a person reads the signals together with the source and the context.
AnyFact’s deepfake video detection follows this pattern. It scores one frame per second with the photo detector, runs a face-swap classifier on the largest face when a frontal face is clearly visible, sends the soundtrack through the voice detector and marks suspicious moments on a timeline. The verdict is probably human, uncertain or probably AI, with the signals and the limits explained.
Detection has hard cases: low-resolution clips, heavy compression, faces in profile and very short snippets. A result is an indication, not proof, and our guide on whether AI detectors are accurate explains why.
Upload the largest, earliest copy you can find. AnyFact accepts MP4, MOV and WebM files up to 200 MB and 3 minutes, or a link to a public video page. For a voice message on its own, the voice detector scores each segment and checks the transcript for scam patterns such as urgency, secrecy, money requests and changed bank details.
AI Video & Deepfake Detector · Free plan
Check a video for deepfake signs: frames, faces and the soundtrack, with suspicious moments on a timeline.
What should you do if you’re targeted by a deepfake?
If you are targeted by a deepfake, preserve the evidence first, then report it to the platform and, where it applies, to the authorities. Acting in that order keeps the proof you may need later.
- 01Save copies: the file if you can, screenshots, the link, the account name and the date you saw it.
- 02Do not engage with the person who posted it, and do not pay anything they ask for.
- 03Report it to the platform using its impersonation, privacy or synthetic media reporting option.
- 04If money was requested or lost, contact your bank at once and report the fraud to the police. In the United States, the FBI asks victims of online fraud to file a complaint at ic3.gov.
- 05If the content is intimate, look at hash-matching services that keep the image on your own device: StopNCII.org, run by the Revenge Porn Helpline at the UK charity SWGfL, for adults, and NCMEC’s Take It Down for images of people under 18. Check each service’s eligibility criteria first.
- 06In the United States, if a covered platform does not remove a reported intimate image within 48 hours of a valid request, the FTC invites people to report the platform at TakeItDown.ftc.gov.
- 07Warn the people who may be contacted next, such as family or colleagues, so they are not fooled.
When you report, be specific. Say that the content is synthetic or manipulated, that it shows you without your consent, and where it appears. Include the links and the dates, and keep a copy of every report you send and every reply you receive.
A detector report that lists its signals can support a platform report or a complaint, but it is not proof on its own. Teams handling these cases will find more on our pages for fraud and KYC teams and for lawyers and legal evidence.
A quick deepfake checklist
Keep this checklist for the moment a clip, a call or a photo feels off. It works with or without a detector.
- 01Pause before you share, pay or reply.
- 02Find the earliest, highest-quality copy of the content.
- 03Look at the face edges, mouth, eyes, hands and light.
- 04Listen for flat emotion, odd pacing and audio that does not match the scene.
- 05Check who posted it first and whether trusted sources confirm it.
- 06Confirm through a channel you already trust.
- 07Run the file through a detector and read the signals, not just the verdict.
- 08Write down what each check found, including the checks that found nothing.
AI Voice Detector · Free plan
Got a suspicious voice message? Score each segment and check the transcript for scam patterns.
Frequently asked questions
Are deepfakes illegal?
Making a deepfake is not illegal in itself in most places, but many uses are, such as fraud, harassment and non-consensual intimate images. Rules differ by country: the EU AI Act requires disclosure of deepfakes from August 2, 2026, and the US TAKE IT DOWN Act targets non-consensual intimate images, including AI-generated ones.
What is the difference between a deepfake and a cheap fake?
A deepfake uses AI to create or alter a face, voice or scene. A cheap fake uses simple edits such as slowing a video down, cutting it or adding a false caption. Both can mislead, so check the source and context in either case.
Can deepfakes be detected?
Often, but not always. Detectors look for traces of generation in frames, faces and audio, and provenance data can name the tool used. Compression, low resolution and short clips make detection harder, so treat any result as an indication, not proof.
How can I protect myself from deepfake scams?
Agree on a secret word or phrase with family, as the FBI recommends, and call back on a number you look up yourself before acting on any urgent request. Be cautious about how much of your voice and image you share publicly.
Is a face filter a deepfake?
Not usually. A beauty or novelty filter alters your own image in a way most viewers recognise. It becomes a deepfake problem when the result is realistic and used to make people believe something false, such as that another person said or did something.
Can AnyFact detect deepfakes?
AnyFact checks videos, voices and photos for signs of AI generation and explains the signals behind each verdict. It has a free plan: a free account gets 10 credits a month, and a video analysis costs 5 credits. The result is an indication, not proof.