How Can You Tell If a Video Is AI-Generated? (And Can Viewers?)

Sometimes you can tell. Often you can't. No single clue, gut feeling, or detector score settles it either way.

That's an awkward answer when you need to decide whether to trust a clip or approve an ad. It's still more useful than another reminder to "check the hands." What works is treating detection as an evidence problem. Collect independent signals, weigh them by how strong they are, and be honest about how sure you actually are.

Why the old checklist stopped working

Most guides on how to detect AI-generated video lead with the same suspects: six fingers, mushy teeth, drifting lip sync, shadows pointing the wrong way.

Those clues helped when generators were worse. Now they fail in both directions.

Current models often get hands, faces and lighting right in any single frame, so a clean still proves nothing. Meanwhile, real footage can look synthetic. Heavy compression smears skin into plastic. Stabilization bends backgrounds. Beauty filters smooth faces and shift jawlines. In low light, the phone's noise reduction turns texture waxy. Fast cuts and speed ramps make motion feel wrong.

So the hand-check approach flags real videos and misses good synthetic ones. It also makes people confident when they shouldn't be.

It helps to know what kind of video you might be looking at. A fully generated video comes from a prompt or an image, with no camera involved. AI-manipulated footage started as real video and then had a face swapped, a voice cloned, or an object added or removed. Ordinary editing, like color grading, retouching and cuts, changes how footage looks without inventing anything. Each leaves different traces, and the same glitch can mean different things in each case. If you want the wider context, here's a primer on synthetic media.

The evidence ladder

Rank evidence by how much weight it can carry and start at the top. Pixel-level clues belong lower than most people put them.

Provenance sits at the top for a reason. Content Credentials are cryptographically signed records describing how a file was made or edited, and a valid one is strong evidence. But metadata often gets stripped when a file is re-exported or uploaded, and a screen recording wipes it entirely. So missing credentials tell you very little. Platform labels work the same way. YouTube may show a "Made with AI" label based on creator disclosure, its own tools, or valid C2PA metadata. A label is useful evidence. No label is not a clean bill of health.

Detectors sit at the bottom. They learn from specific generators and tend to struggle with newer ones, heavily compressed files, and clips that have been re-encoded several times. Treat a score as a reason to look harder, never as a verdict.

How to review a suspect video: pause, replay, verify

This is how careful reviewers work in practice. It takes a few minutes per clip.

1. Watch once at normal speed, with sound. Don't hunt yet. Note where your attention snagged: a moment that felt off, a voice that didn't sit right, a product that looked wrong. First impressions are weak evidence but good pointers.
2. Replay the moments that snagged, two or three times, at half speed.
3. Scrub frame by frame through those sections. You're looking for things that change when they shouldn't.
4. Check the details that have to stay stable over time:
   • Text, logos and labels should stay identical as objects move. Watch for letters that warp, shimmer or change spelling mid-shot.
   • Products should keep their geometry. A phone shouldn't gain a camera lens, and a bottle shouldn't change proportions as it turns.
   • Objects should persist. Watch what happens when a hand passes in front of something. Does it vanish, or come back slightly different?
   • Contact and physics should hold up. Fingers should press into what they grip, liquid should pour and settle like liquid, and hair should move with the head.
   • Faces should stay the same person. Earrings, moles, teeth and hairlines should survive angle changes and cuts.
   • Reflections and shadows should track the objects casting them through the whole shot.
   • Lip sync should hold for the full clip, not only the first few seconds, and room sound should match the space on screen.
5. Verify outside the frame. Find the earliest upload. Check the account's history. Look for other footage of the same event from other angles and other sources. Check for Content Credentials or a platform label.

Then weigh what you found. A problem that persists or repeats across frames counts for far more than one odd still. One warped finger in one frame of a compressed clip is noise. A logo that changes spelling three times in four seconds is a signal. Generators lose track of people and objects over time for predictable reasons, which our guide to AI video character consistency covers in detail.

"Is it AI?" and "Is it true?" are separate questions

People tend to merge these two questions, and it causes bad calls.

The first asks whether the video was generated or manipulated. The second asks whether the event, claim or person in it is real and trustworthy. They overlap, but answering one doesn't answer the other.

Real footage can mislead when it comes from another place or another time, or when the context has been cut away. AI-generated video can be harmless: a clearly stylized ad, an explainer, a parody. And provenance only records how a file was made. It doesn't certify that what it shows actually happened.

For high-stakes content, like a claim about a company, a public figure, a safety issue or a financial decision, source checks and corroboration matter far more than pixel-hunting. If you can't trace where a video came from and nobody else has footage of the event, that's a bigger problem than whether the shadows look right.

So, can you tell if a video is AI-generated just by watching?

Not reliably. Research keeps landing on that answer, though the numbers vary a lot depending on how each study was set up.

In a pre-registered survey series with 1,276 participants judging images, audio, video and audiovisual clips, mean accuracy across all stimuli was 51.2%. Video-only clips came in at 50.7% and audiovisual clips at 54.5%. That's close to a coin toss.

A different study had 490 US participants each judge 13 videos drawn from a varied set of real-world clips. Mean accuracy was 66.3%. People identified authentic videos 75.6% of the time but manipulated ones only 59.4% of the time. Confidence was often poorly matched to accuracy, and 91.5% of participants didn't know about analysis or verification tools.

You can't compare those numbers directly, because the clips, tasks and generators were different. The pattern still holds across both: people pick up some cues, they're inconsistent, and they do worse at catching manipulated video than at recognizing real footage. Feeling sure is not evidence.

For advertisers, this cuts both ways. Most viewers won't reliably identify an ad as AI-made. But they don't need to name it to feel that something is off. A product that subtly changes shape, or a spokesperson whose teeth shift between shots, can come across as cheap or untrustworthy before anyone thinks the word "AI."

What this means for AI video ads

If you're buying or approving AI video ads, the question changes. You already know the clip is AI. What you need to know is which visible failure will make the ad feel unreliable or pull attention away from the offer.

An ad doesn't need to pass as camera footage. Stylization can be a deliberate choice, and a clearly illustrated or heightened look can work well. Failed realism is what hurts: footage that tries to look filmed and almost gets there. Viewers notice the "almost."

Spend review time where the buyer's attention goes:

• The product has to match reality in every frame: label, logo, colors, shape, packaging. A misspelled brand name on screen for a third of a second is enough to sink the ad, and in some categories it becomes a compliance issue.
• The spokesperson should stay the same person across cuts, with the same face, voice, outfit and accessories. Identity drift reads as fake faster than almost anything else.
• Any moment where someone holds, opens or uses the product has to look physically right. That moment is where the offer lives.
• On-screen demonstrations of results, like before-and-after shots, measurements or a product doing its job, have to be honest. A generated demo that overstates what the product does is a trust problem and possibly a legal one, however good it looks.
• Disclosure should be set correctly for the platform. Check whether it requires a label for realistic AI content before the ad goes live.

Background oddities in a stylized scene may be fine. A wobbling logo on your hero product is not.

Most of this should be caught before you ever see a cut. Strong creators review their own output frame by frame, regenerate bad shots, and never deliver a melting label. That's the practical case for keeping a skilled human in the loop. If you'd rather hand that responsibility to someone who does it daily, here's what to look for when you hire an AI video creator.

Quick checklist: pause, replay, verify

• Watch once at normal speed with sound and note where your attention snagged
• Replay those moments at reduced speed
• Scrub frame by frame through suspect sections
• Check text, logos and product shape over time
• Check object permanence, contact and physics
• Check face and accessory continuity across angles and cuts
• Check reflections, shadows and lip sync across the full clip
• Rule out compression, stabilization, filters, low light and heavy editing as the cause
• Look for Content Credentials or a platform AI label
• Trace the original source and look for corroborating footage
• Treat any detector score as one input, if you use one at all

The decision rule

One weird frame is suspicion. Several independent signals plus weak provenance justify escalation, not certainty.

Escalation means you hold the share or the approval, contact the source, and look for corroboration. Certainty only comes from strong provenance or solid outside verification, and sometimes you won't get either. When that happens, say so plainly instead of guessing.

For ads, the rule is simpler. If a visible failure touches the product, the spokesperson or the claim, fix it before the ad runs. Everything else is a call about style.