How to Tell if a Video or Image Is a Deepfake

The clip looks real, your gut says wrong, and you're about to hit share. Here's how to slow down and actually check before a fake spends your trust for you.

A clip lands in your group chat. A local official says something ugly, or a celebrity endorses a coin, or a "leaked" security camera shows your neighbor doing something unforgivable. Your stomach drops. Your thumb is already over the share button, because outrage moves faster than doubt. That half-second — the gap between feeling it and checking it — is exactly where a deepfake does its work. It doesn't need you to believe it forever. It just needs you to react before you verify.

The good news: you don't need forensic software to be hard to fool. You need a habit. Here's the NU version of that habit — what to look at, what to trust, and why staying skeptical without going crazy is the whole game.

Look at the seams, not the face

Generative models are great at faces and bad at everything around them. So check the edges of the illusion:

• Hands, ears, teeth, and jewelry. Count fingers. Look for earrings that change shape, teeth that smear when someone talks, glasses whose arms don't connect to the frame.
• Edges and hair. Where a face meets a neck, or hair meets background, look for shimmer, warping, or a too-clean cutout — like a sticker laid on the scene.
• Lighting and reflections. Does the light on the face match the room? Are shadows pointing the same way? Eyes are tiny mirrors; real ones often catch the same light source.
• Mouth-to-audio sync. In lip-synced fakes the mouth interior looks rubbery, and consonants like "b," "p," and "m" don't fully close the lips.
• Blinking, breathing, micro-movement. Stillness that's too still, or motion that loops oddly, is a tell.

A warning that matters: these tells are disappearing. Every month, models fix last month's giveaways. Treat "I didn't spot anything weird" as not proof of real — only as "the visual layer didn't catch it." That's why the next two steps carry more weight than your eyes.

Chase the provenance: where did this actually come from?

The single most useful question is not "does it look fake?" but "where did this file originate, and who first posted it?"

• Find the origin, not the repost. A screen-recorded clip with a caption is evidence of nothing. Look for the first, highest-resolution version from an account or outlet that can be held accountable. If the only sources are anonymous reposts, that's a flag.
• Check for content credentials (C2PA). C2PA is an open industry standard backed by Adobe, Microsoft, and major camera and chip makers. It attaches a tamper-evident record to a file — what device or tool made it, and what edits happened. Some cameras and AI generators now stamp it automatically. You can inspect it at the Content Authenticity Initiative's free Verify tool (contentcredentials.org/verify) by uploading the file or its link.
• Know C2PA's limits. Provenance is powerful but not magic. Most files in the wild carry no credentials at all, and a screenshot or re-encode can strip them. So: credentials present and intact = strong signal it's genuine; credentials absent = unknown, not guilty. Absence proves nothing either way.

Reverse-search the image (and the video's frames)

Most viral fakes aren't built from scratch — they're recycled or recontextualized real media.

• Reverse image search the picture or a clear video frame using Google Lens, TinEye, or Bing Visual Search. You're looking for the same image showing up earlier, in a different place, or with a different caption. A 2019 photo relabeled as today's disaster is one of the oldest tricks alive.
• Grab keyframes from a video (pause and screenshot a sharp frame) and search those. Old footage re-dressed as breaking news collapses fast under this.
• Read the metadata when you can. Tools like the InVID/WeVerify browser plugin help pull video keyframes and metadata. Treat metadata as a clue, not a verdict — it's easy to fake or wipe.

Use detectors as a vote, never a verdict

AI "deepfake detectors" exist, and some are decent. They are also frequently wrong in both directions, and they degrade the moment a new generator ships. Use them the way you'd use one witness: as input, not as the judge. If a detector, the provenance check, and a reverse search all point the same way, you have something. If they disagree, you have a "don't know" — and "don't know" is a legitimate, honest answer to publish.

Why "kooky till proven" cuts both ways

NU's posture is kooky till proven — don't dismiss a strange claim out of hand, but don't crown it true until the record backs it. Deepfakes weaponize both failure modes. Hoaxers exploit the people who believe instantly. But there's a quieter danger called the liar's dividend: once everyone knows video can be faked, real footage gets waved away as "probably AI." Genuine evidence of genuine wrongdoing gets dismissed because doubt became a reflex.

So the discipline isn't "trust nothing." It's "hold the claim open, then go get the record." A thing can be unverified and still be true; it can also be vivid, emotional, and completely fabricated.

Before you share, run the checklist

• Who first posted this, and can they be held accountable?
• Is there a higher-res original, or only reposts and screen-records?
• Does it carry intact content credentials (check contentcredentials.org/verify)?
• Does a reverse search show the image or frames appearing earlier or elsewhere?
• Do the physical tells (hands, light, lip-sync) hold up?
• Does a second source confirm it independently?

If you can't answer most of these, the honest move is to wait, not amplify. A fake only spends the trust you hand it. Make it earn yours.