How to Spot a Deepfake Video

How can you tell if a video is AI-generated or a deepfake?

You tell if a video is AI-generated by checking three layers: the picture, the sound and the story around the clip. Knowing how to spot a deepfake video comes down to finding a mismatch in at least one of them, then confirming it with the others.

Do not judge at normal speed on a phone. Watch on a large screen, pause, step frame by frame and slow playback down. Most tells last a fraction of a second.

Face swap, lip sync or fully generated: which kind of fake is it?

There are three main kinds of AI video fake, and each one fails in different places. Naming the likely type tells you where to look.

Keep the last row in mind. Many misleading clips are real footage with a false caption, and no fake video detector will flag them. Only source checks catch them.

Fully generated clips tend to be short, with a single continuous shot or a few loosely connected ones. Face swaps and lip-sync edits, by contrast, are often built on long real footage such as interviews, speeches or webcam recordings. The length and format of a clip are therefore a first hint about which checks to run.

What visual tells reveal a deepfake?

The most revealing visual tells appear where the fake must blend with the real: the border of the face, the inside of the mouth and the way light falls across moving features.

What audio tells reveal a fake video?

The soundtrack often gives a fake away because cloned voices and real footage rarely share the same acoustics. Close your eyes and listen to the clip on its own.

• The voice sounds too clean for the room, with no echo in a hall or no wind outdoors.
• Breathing is missing, or breaths fall in odd places.
• Intonation is flat or oddly regular, with emotion that does not match the face.
• The background noise cuts out when the person speaks, then returns.
• Words land slightly before or after the mouth moves.

Be suspicious of background music laid over a speech. Music can be a creative choice, but it also masks the artifacts and acoustic mismatches that would give a cloned voice away. If you can, find a version of the clip without the music.

Our guide on how to tell if a voice is AI goes deeper on cloned speech and the scam scripts that often come with it.

How do you check the source, date and context of a video?

You check the context by finding the first upload and comparing it with what the clip claims to show. This is the single most effective way to detect fake video, because it also catches real footage presented out of context.

1. 01Take screenshots of three or four key frames: a clear face, a wide shot, any visible sign or landmark.
2. 02Run each frame through a reverse image search and sort results by date.
3. 03Find the earliest upload and the account that posted it. Is it the person shown, a known outlet or an anonymous account?
4. 04Compare the upload date with the event the clip claims to show.
5. 05Look for the same moment from other angles, other witnesses or a full-length version.
6. 06Ask whether the speaker would plausibly say this, here, in this way.

Frames from a video are images, so the methods in our guide on how to tell if an image is AI generated apply to each key frame too.

Do platforms label AI-generated videos?

Some platforms label AI-generated videos, but you cannot rely on the label being there. Labels depend on the creator disclosing it or on the platform finding provenance data in the file.

A label that says a clip was made or edited with AI is useful evidence. A missing label proves nothing: uploads are re-encoded, metadata is stripped, and many tools add none to begin with. Learn what that provenance data is in our guide to C2PA and Content Credentials.

Read labels carefully. “Made with AI” and “edited with AI” can cover anything from a fully generated scene to a cleaned-up soundtrack or a replaced sky. Look at the description and the creator’s other posts to understand what was actually changed. Watermarks and logos from video generators help too, but they are easy to crop or blur, so their absence means little.

How do you use a deepfake detector and read its timeline?

You use a deepfake detector to find the moments worth inspecting, then you check those moments yourself. A good AI video detector tells you when it saw something, not just whether.

The AnyFact deepfake video detector accepts MP4, MOV and WebM files up to 200 MB and 3 minutes, or a link to a public video page. 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, and sends the soundtrack through the voice detector.

Suspicious moments are marked on a timeline. Jump to each one, step through it frame by frame and compare with the tells above. In the grey zone, Claude reviews the five most suspicious frames as a second opinion; it can move the score only a little and never decides alone.

How do you spot a deepfake on a live video call?

You spot a deepfake on a live call by making the other person do things that real-time face and voice swaps handle badly, and by verifying them outside the call. Real-time fakes have less time to render each frame than a pre-recorded clip, so they tend to break under movement and occlusion.

• Ask them to turn their head slowly to a full profile, then back.
• Ask them to pass a hand in front of their face, or to touch their nose or ear.
• Ask them to stand up, move closer to the camera or change the lighting in the room.
• Ask a question only the real person could answer, about something not posted online.
• Watch for a frozen or blurred face that recovers when they stop moving.

None of these tests is decisive, and a poor connection causes glitches too. The decisive step is the same as for phone scams: if the call leads to a payment, a password or confidential data, end it and contact the person through a channel you already trust before doing anything.

How to spot a deepfake video: a 9-step checklist

1. 01Decide which kind of fake is plausible: face swap, lip sync, fully generated or out-of-context real footage.
2. 02Find the original upload, its account and its date.
3. 03Reverse-search three or four key frames.
4. 04Watch the face edges, mouth, teeth and eyes in slow motion.
5. 05Check lighting, physics, hands, background people and text.
6. 06Listen to the audio alone for acoustics, breathing and timing.
7. 07Look for a platform label or Content Credentials.
8. 08Run the clip through a deepfake detector and inspect each flagged moment.
9. 09If the evidence is mixed, call the video unverified and do not share it as fact.

Teams that review video at scale will find workflows for trust and safety moderation and for journalists and fact-checkers.

Frequently asked questions