How to Tell If a Video Is AI-Generated: Signals That Actually Matter

Visual tells still matter. They just work better as part of a pattern than as a magic trick.

• Unstable fine detail: Hair strands, fabric texture, eyelashes, jewelry, or lettering may sharpen and soften in inconsistent ways from frame to frame.
• Lighting that feels locally wrong: Faces and objects can carry highlights or shadow edges that do not agree with the scene’s broader light direction.
• Edge tension around composites: Jawlines, glasses, fingers, microphones, and moving foreground objects sometimes show haloing, soft seams, or odd boundaries.
• Background logic breaks: Crowds, signage, screens, or room details may morph subtly when the subject moves even though the camera perspective has not changed enough to justify it.

This is where frame grabs fall short. A paused image can look ordinary while the sequence behaves unnaturally.

• Flicker under motion: Facial detail, clothing texture, or object edges pulse as the subject turns, speaks, or crosses the frame.
• Motion drift: Object geometry shifts slightly during movement, especially in shoulders, teeth, hands, or the boundary between face and background.
• Landmark instability: Eyes, mouth corners, nostrils, or jaw geometry drift more than the head movement should produce.
• Mouth-shape mismatch: Visible mouth positions do not fully line up with the phonemes or cadence that the audio suggests.
• Scene persistence errors: A detail that should remain stable through a shot changes identity, such as earrings, lapels, wall objects, or the outline of a chair.

Reviewers often focus on pixels first and provenance later. In higher-stakes cases, that order is backwards.

• Export and packaging traces: Codec, container, and timing fields can suggest platform re-exports, editing steps, or other packaging history.
• Missing or contradictory metadata: Absent fields are not proof of manipulation, but inconsistent timestamps or packaging details are worth documenting.
• Provenance gaps or signed credentials: If a clip has no credible chain back to a source capture, your claim should stay limited. If valid credentials exist, they materially improve context.

If you are new to provenance review, our explainer on content credentials and C2PA is a good companion.

1. 01
   Preserve the best available copy

   Avoid making decisions from a screenshot or a repost if the original file or closest-to-original export can be obtained.

2. 02
   Review signal categories separately

   Inspect visual, temporal, facial, audio, metadata, and provenance evidence as distinct streams before combining them.

3. 03
   Record evidence quality

   Note when audio is missing, the clip is very short, or the file has been re-encoded enough to reduce confidence.

4. 04
   Check source context and provenance

   Ask where the clip came from, how it was obtained, and whether a source, platform, or creator can provide corroboration.

5. 05
   Escalate when the consequences are high

   If the clip affects publication, moderation, public safety, or evidence handling, move it to human review instead of forcing certainty from automation.

Human review becomes especially important when the stakes are legal, reputational, or fast-moving; when the clip has weak provenance; when the media is short or degraded; or when the result is strong but the surrounding source story still does not add up.

A high AI-likelihood score is not proof. A low AI-likelihood score is not an authenticity certificate. Strong claims still need source validation, provenance review, and human judgment.