看到疑似AI生成的视频，应该先看画面破绽吗？ / Should You Look for Visual Glitches First in a Suspected Deepfake?

Judgment in the Age of AI · Article 16

Lip movement, blinking and inconsistent lighting were once common clues to deepfakes, but improving generation makes visual glitch-hunting increasingly unreliable. A stronger order verifies the distribution context before examining pixels.

Layer one: the source chain

• Who first published it, rather than merely reposted it?
• Does the account belong to the person, organisation or a credible news outlet?
• Can the complete material be found through an official channel?
• Are the beginning, ending or surrounding remarks missing?

Layer two: event consistency

Check whether place, time, clothing, event arrangements and public schedule align. A technically perfect fake can still fail the logic of the event. Reverse-searching key frames may reveal old footage that has been cropped or redubbed.

Layer three: file and image

Only then inspect audio synchronisation, blurred edges, skin changes, background jumps and unnatural motion, together with platform labels or provenance credentials. No single glitch proves fabrication, and the absence of glitches does not prove authenticity.

The default action before sharing

If a video could harm a person, affect a public event or demand immediate action, stop distribution and seek independent confirmation.

References

• eSafety Commissioner: Deepfakes
• Scamwatch: How Scammers Use Technology and AI