What is a deepfake? How it works, risks and detection

A deepfake is an image, video, or audio recording that has been created or altered with artificial intelligence so it appears authentic. A deepfake may replace a face, change an expression, clone a voice, or generate a person or event that never existed. The term usually describes realistic synthetic media involving a person's identity.

Deepfakes are not automatically malicious. They can be used with permission for entertainment, translation, accessibility, education, and visual effects. The harm comes from deception or unauthorized use: impersonation fraud, non-consensual intimate imagery, fabricated evidence, and false political or commercial messages.

Deepfake meaning in one minute

The U.S. Government Accountability Office defines deepfakes as videos, audio, or images that seem real but have been manipulated using AI. That definition covers several techniques:

• Face swapping: placing one person's facial identity onto another person's photo or video.
• Face reenactment: changing a person's expression or mouth movement so they appear to perform an action they did not record.
• Voice cloning: synthesizing speech that resembles a real person's voice.
• Full synthesis: generating a person, scene, or event rather than altering a specific original recording.

Not every AI-generated image is a deepfake. A clearly fictional landscape made from a prompt does not impersonate a real person. A conventional edit that adjusts exposure is also not ordinarily called a deepfake. The useful questions are whether AI changed the apparent identity, speech, action, or origin of the media—and whether a viewer could reasonably mistake it for a real record.

How do deepfakes work?

A system learns patterns from examples, then generates or transforms media to match a target. In a face swap, the source provides the identity and the target photo or video provides the pose, expression, lighting, and scene. The system estimates facial landmarks and produces replacement pixels across the sequence.

Video is harder than a single photo because the result must remain consistent as a person turns, speaks, passes behind an object, or moves through different lighting. Errors often appear around hair, teeth, glasses, hands crossing the face, or frames with motion blur.

Voice cloning follows a related idea in audio: a model learns characteristics of a voice, then synthesizes new speech. Modern impersonation attempts may combine generated voice, an altered profile image, and ordinary social engineering. The technical artifact is only one part of the deception.

Older deepfake explanations often focus on generative adversarial networks, or GANs. Current synthetic-media systems can also use diffusion and other model architectures. The practical definition should focus on the result and its use rather than assuming every deepfake was made with one specific model family.

Are face swaps and deepfakes the same thing?

A face swap is one technique that can produce a deepfake, but the terms are not interchangeable. A clearly labeled, consensual face swap for a film effect is still synthetic media, yet it may not be intended to deceive. A deepfake can also be made without swapping a face, such as by cloning a voice or generating a false video of a person speaking.

If your goal is an authorized creative edit, see how to swap faces in a photo or video. For choosing a workflow, compare AI face swap tools by photo, video, multiple-face, local-processing, and export requirements.

How can you spot a deepfake?

Do not rely on one visual glitch or a single detector score. The GAO's 2024 review says current detection methods have limited effectiveness in real-world conditions and may perform worse when the lighting, expressions, media quality, or generation method differs from their training data.

Use this verification sequence instead:

1. Check the source. Find the earliest available post, the publisher's account, and the surrounding page. A familiar logo or username can be copied.
2. Confirm the claim elsewhere. For an alleged speech, transaction, emergency, or public event, look for an independent first-party statement or credible reporting.
3. Verify the person through a known channel. If a caller asks for money, credentials, a code, or urgent secrecy, end the exchange and contact the person using a number or account you already trust.
4. Inspect the complete media. Watch for unstable facial boundaries, inconsistent reflections, irregular accessories, implausible shadows, lip-sync errors, unnatural motion, or a voice that does not match the person's normal phrasing. These are reasons to investigate, not proof on their own.
5. Look for provenance. Original files, publication history, signed content credentials, or an attributable production record can provide stronger evidence than visual inspection alone. Missing metadata is not proof of fabrication because platforms routinely remove it.
6. Use detection tools as supporting evidence. Record the tool, version, file, and score. Do not turn one classifier result into a categorical verdict.

The FBI's guidance on AI-generated impersonation recommends independently verifying the sender, inspecting contact details and links, and refusing requests for money or sensitive information until identity is confirmed. Those checks remain useful even when the media contains no obvious defect.

What should you do if a suspected deepfake targets you?

• Preserve the original message, URL, username, timestamp, file, and screenshots before the content changes.
• Do not negotiate with an impersonator or send money to make the content disappear.
• Warn affected contacts through an account or channel you control.
• Report the content and impersonating account to the platform where it appears.
• For fraud, threats, intimate-image abuse, or immediate danger, contact the appropriate local authority or qualified support organization.
• Change compromised passwords and enable multifactor authentication if the incident may involve an account takeover.

Avoid repeatedly reposting the suspected media as a warning. Copies can extend its reach and make the false claim harder to contain. Link to a correction or official statement where possible.

Legitimate uses of synthetic media

Permission and context distinguish many legitimate uses from abuse. Examples include an actor approving a translated performance, a brand creating a disclosed visual effect, a teacher illustrating media literacy, or a filmmaker replacing a stunt performer's face under a production agreement.

A responsible workflow should answer four questions before generation:

• Do you have permission to use the person's face, voice, and source media for this purpose?
• Will the result make a real person appear to endorse, say, or do something they did not approve?
• Does the audience need a label or other context to understand that the media is synthetic?
• Can you preserve the inputs, consent record, approved output, and publication context?

Consent to take a photo is not automatically consent to train a model, clone a voice, create intimate content, or use a likeness in an advertisement. Requirements vary by jurisdiction and platform, so check the rules that apply to the intended use.

How to make an authorized face-swap video

Use footage and likenesses you have permission to edit. Choose a clear replacement portrait, upload the target video, map each person deliberately, and review the complete result before sharing it. A short, well-lit clip with a mostly visible face is easier to inspect than a long sequence with frequent cuts.

Magic Hour provides separate workflows for single-person video face swap and multiple-face video swap. A successful generation does not establish that you have the right to publish it; permission and context remain the creator's responsibility.

Deepfake questions people ask

Is a deepfake always a video?

No. The term can refer to manipulated or synthetic images, audio, and video. A cloned voice used in an impersonation attempt is synthetic media even when no video is involved.

Can you detect a deepfake just by looking at it?

Sometimes visible errors raise suspicion, but appearance alone is unreliable. Generation methods improve, compression can hide artifacts, and authentic recordings can also look unusual. Source verification, independent confirmation, and provenance provide a stronger process.

Are all face swaps illegal?

No universal answer applies to every jurisdiction and use. Permission, deception, publicity rights, privacy, sexual content, fraud, defamation, and local synthetic-media rules can all matter. A consensual creative effect is different from impersonation or non-consensual content.

What is the difference between a deepfake and a cheapfake?

A deepfake uses AI to create or alter media. A cheapfake uses simpler methods such as cropping, slowing, speeding, relabeling, or presenting real footage in a false context. Both can mislead, which is why verifying the source and claim matters more than identifying one production technique.

Can a deepfake detector prove that media is fake?

Treat detector output as one piece of evidence. The GAO notes that real-world accuracy can fall when inputs differ from the detector's training data, and generation techniques continue to change. Important decisions require corroboration and, where appropriate, forensic review.

The practical takeaway

A deepfake is convincing AI-generated or AI-altered media involving apparent identity, speech, or events. To evaluate one, verify the source and claim, contact the person through a trusted channel, inspect the complete file, and use detection tools only as supporting evidence. To create synthetic media responsibly, obtain permission, preserve context, and never turn an authorized creative tool into an undisclosed endorsement or false record.