AI-Generated Video: 7 Clues to Review and Their Limits (2026)
An odd blink or a blurred face can be a reason to inspect a video more closely. It cannot establish that the video was made with AI. Compression, filters, editing and recording conditions can produce similar effects, while generated videos may show none of these clues.
Use the seven checks below to record what you see and where it happens. Then compare those observations with the source, the original file and other reporting.
1. Blinking and eye movement
Watch whether the eyes and eyelids remain consistent when the person turns or looks away. Note an eyelid that changes shape or a gaze that jumps between nearby frames.
Do not use a fixed blinks-per-minute cutoff. People blink differently when reading, speaking or concentrating, and a short clip may contain no blink. A natural blink pattern also does not establish authenticity.
2. Face edges, skin and lighting
Look around the jaw, ears, hair and glasses as the head moves. A face boundary that separates from the head or changes independently of nearby objects deserves review across several frames.
Smooth skin can come from beauty filters, low resolution or noise reduction. Changes in lighting can come from exposure adjustment or a moving light source. Compare the suspected face with the rest of the scene before interpreting the effect.
3. Background and object continuity
Follow signs, door frames, jewelry, reflections and objects as the camera moves. Record a specific change, such as lettering that changes between adjacent frames or an object that disappears without being hidden.
Motion blur, stabilization, transitions and platform compression can distort those details too. Check the best available copy and replay the surrounding sequence; one distorted frame is weak evidence.
4. Lip movement and speech timing
Play the clip at normal speed first, then inspect a short section around a suspected mismatch. Check whether the audio stays out of sync or whether the issue occurs only at a cut.
Dubbing, streaming delay, an edited soundtrack and playback problems can all cause a mismatch. Good synchronization does not rule out synthesis. For a claimed speaker, verify the source of the recording separately.
5. Voice and recording texture
Listen for abrupt changes in tone, repeated segments or a voice that changes unexpectedly within a sentence. Compare with a known recording only when the speaker and recording conditions are reasonably comparable.
A smooth voice, few pauses or little breath noise is not proof of cloning. Studio processing and noise suppression can produce those qualities, and synthetic speech can include pauses and breaths. A familiar voice is also insufficient to authorize a payment or disclose information.
6. Teeth and mouth details
When the mouth is visible, compare its shape and teeth across nearby frames. Look for details that change without a corresponding movement, rather than judging whether a single paused frame looks strange.
Small teeth and tongue details are easily blurred by compression, motion and low light. Avoid treating a blurry mouth as a reliable test, especially in a downloaded social-media copy.
7. Hands and interactions
If hands are visible, follow their contact with objects and their position over time. A finger changing shape or an object passing through a hand merits a closer check of the surrounding frames.
Hands being off-screen, cropped or blurred is not evidence of AI generation. Fast movement, occlusion and ordinary framing can explain those conditions. Count visible fingers only when the image is clear enough to support the observation.
What to do with a suspicious clue
Preserve the post URL, caption and original file where possible. Search several representative frames for earlier versions. Check the date, location and claimed event against independent sources. For impersonation or an urgent financial request, contact the person through a channel you already trust.
Review any Content Credentials, but keep their scope clear: a valid provenance record does not verify the truth of the event, and missing credentials do not prove a fake.
You can use the video detector as an additional check. DeepFakeCheck accepts MP4, MOV and WebM files up to 100 MB and samples up to 10 timestamped frames. It does not inspect every frame and may miss a brief edit between samples. A low-risk result cannot authenticate the speaker or story; a high-risk result still needs corroboration.
A limited free allowance is available without signup. Analysis content is retained for up to 7 days, then deleted. Processing time varies with the file and service conditions.
For the complete source-checking sequence, read How to Detect a Deepfake Video. If the evidence remains inconclusive, keep that uncertainty in any description you share.
Sources and further checks
The following resources support the verification workflow; they do not establish a detection rate for these seven visual checks.