How to Detect a Deepfake Video: 8 Practical Checks That Still Work

How to Detect a Deepfake Video

The deepfake generators are getting better every year. The detection tools are always a step behind. What you can do is build a habit — a quick, repeatable check that catches most synthetic clips and surfaces a "this needs a second look" flag for the rest. The eight checks below are the ones that hold up on real-world output.

The bar is "needs a second look," not "this is fake"

The output of these checks is rarely a clean verdict. It is usually a confidence level: high, medium, or low that the video is synthetic. Use the checks to triage. Use the provenance test — find the original — to confirm.

Warning

Check 1: Watch the face boundary

The single most reliable tell on a current-generation deepfake is the seam where the generated face meets the real background. Watch the boundary of the face at high zoom and at full speed.

What to look for:

• A faint shimmer or color mismatch along the jawline
• Hair strands that appear and disappear at the edge of the face
• Earrings that change shape across frames
• Glasses frames that do not match the angle of the face
• A neck shadow that does not quite match the rest of the lighting

These are subtle. They are not a verdict on their own. Combined with another check, they are damning.

Check 2: Watch the eyes and the blink rate

Early deepfakes rarely blinked. Newer ones blink on a normal cadence. The current generation sometimes over-blinks, or blinks out of sync with the speech.

What to look for:

• A blink rate that is unusually high or low for the context
• Blinks that do not line up with the cadence of the speech
• Eyes that never fully close, or close too slowly
• Pupils that change size inconsistently with the lighting
• Eye contact that is unusually steady, or unusually shifty

Eye tells are not what they were three years ago, but they have not gone away.

Check 3: Listen to the audio and watch the mouth

A real human speaks with breath, hesitation, and small mouth movements that the audio and the video agree on. A deepfake can produce clean audio, or audio that does not match the lip movements, or both.

What to look for:

• Audio that is unusually clean — no background noise, no room tone, no breath sounds
• Lip movements that do not quite match the audio
• A voice that has a slightly metallic or smoothed quality, especially on sibilants (s, sh, ch sounds)
• A long speech that has no breaths, or has breaths at unusual points

The audio test is one of the strongest. A clean audio track on a video that claims to be filmed in a noisy environment is a major tell.

Check 4: Split the video into frames

Most deepfake tells only show up at full resolution on a single frame. Tools like InVID / WeVerify and a free ffmpeg command will let you pull keyframes from the video and look at them individually.

The workflow:

1. Pull 20-30 keyframes spaced across the video.
2. Look at each one at full resolution.
3. Compare the same facial feature across frames — a watch face, a piece of jewelry, a small reflection.
4. If the feature changes inconsistently, the video is suspicious.

A real video has physics. A deepfake has plausible-looking pixels that are usually consistent within a frame but inconsistent across frames.

Check 5: Look at the hands and the body

Hand synthesis is improving, but it is still less reliable than face synthesis. In a video that shows a person from the chest up, the hands (if visible) are often the tell.

What to look for:

• Hands that appear and disappear at the edges of the frame
• Fingers that merge or have the wrong number of knuckles
• A watch face that changes shape across frames
• A wedding ring that switches hands
• Clothing details that change shape across the video (buttons, collar, zipper)

A video that focuses tightly on the face and shows no hands is a partial workaround. A video that shows hands, watch, jewelry, or detailed clothing is a richer target for tells.

Check 6: Check the source

A deepfake video is often shared with a thin source — a single social account, no original posting, no archival record. A real video, especially one of a public event, usually has multiple sources and a clear origin.

What to look for:

• Who posted the video first, and when?
• Is there an earlier version of the same video, with a different quality or a different angle?
• Does the source have a track record of original footage, or do they only share viral content?
• Is the event covered by other outlets, with their own footage?

A video with a thin source and a strong claim is a yellow flag. A video with a thick source — multiple witnesses, multiple angles, archival record — is much harder to forge convincingly.

Check 7: Reverse image search the keyframes

Pull a distinctive frame from the video and run it through TinEye, Google Images, and Yandex. You are looking for the same frame appearing elsewhere, ideally on a different platform or with a different date.

• If the frame appears in older posts, the video is recycled.
• If the frame appears in higher resolution, the video is a re-encode of an earlier video.
• If the frame does not appear anywhere, the video is fresh — which is a yellow flag for a strong-claim video, but not a verdict.

Check 8: Look at the metadata

A real video carries metadata — the device that recorded it, the date, the GPS if location services were on. A deepfake often has its metadata stripped by the re-encoding that happens when it is uploaded to a social platform.

A video with rich metadata (real device, real date, real GPS) is a small positive signal. A video with no metadata is not a verdict — most re-shared videos have no metadata — but a video with contradictory metadata is more interesting. A claimed iPhone video with EXIF from a 2014 Android device is a finding.

The two checks that do not work

A few common moves are no longer reliable:

• "The eyes are slightly off." Better generators now render eyes well enough that this is not a useful tell on its own. Use it combined with another check.
• "There is no metadata." Every social platform strips metadata. Absence of metadata is the default, not a tell.

The test that survives the next generator

When the visual and audio tells are mixed, do the only test that actually scales: find the original.

A real video of a real event was originally posted somewhere, by someone who was there. A synthetic video of a real event was generated by someone who was not. The provenance test — find the earliest version, in the wild, taken by a person with a track record — is the only test that has not been beaten by the next generation of generators. It is also the test that takes the most time, which is why you should reach for it only when the other checks leave you uncertain.