How to Tell If a Video Is AI Generated: 9 Signs to Check

A clip lands in your feed. A politician says something outrageous, or a celebrity endorses a coin you have never heard of. It looks real. It sounds real. But something feels off, and you want to know how to tell if a video is AI generated before you share it.

A clip lands in your feed. A politician says something outrageous, or a celebrity endorses a coin you have never heard of. It looks real. It sounds real. But something feels off, and you want to know how to tell if a video is AI generated before you share it.

Here is the honest starting point: AI video quality jumped dramatically through 2025 and 2026. Sora, Runway, and Veo now produce footage that fools most people at a glance. Your eyes alone are no longer a reliable test, but they are still the right first step.

Quick Answer: The 9 Signs a Video Is AI Generated

Scan a suspicious clip for these nine signs, then verify with a detector:

1. Lip-sync drift and mouth shapes that do not match the audio
2. Unnatural blinking and frozen or mismatched eye reflections
3. Warped hands, fingers, and jewelry
4. Physics errors such as floating objects, clipping, or slow falls
5. Backgrounds that morph between frames
6. Lighting and shadow mismatches
7. Garbled text, logos, and signage
8. Audio that does not match the scene
9. Too-smooth skin and repeating textures

No single sign is proof. Two or more together is a strong signal worth verifying.

How to Tell If a Video Is AI Generated by Eye

Manual checks still catch low and mid-quality fakes, which make up most of what spreads on social media. Work through these nine signs in order. Each one targets a specific weakness in how video generators build frames.

1. Lip-Sync Drift and Mouth Shape Errors

Watch the mouth during plosive sounds: p, b, and m. A real speaker's lips close completely and in perfect time with the audio. AI-generated and lip-synced video often runs one to three frames off, and the inside of the mouth is a common failure zone. Look for teeth that blur or merge into a single white band, and a tongue that never quite moves naturally. This sign remains useful even against newer models, because precise audio-to-mouth alignment is still hard to fake.

2. Unnatural Blinking and Frozen Eye Reflections

Early deepfakes barely blinked. Modern generators blink, but often at eerily regular intervals, like a metronome. The better tell now is reflections. Real eyes mirror the light sources in the room, and both eyes should show the same reflection. AI video frequently renders mismatched catchlights, or reflections that stay frozen while the head turns. Pause on a close-up and compare the two eyes directly.

3. Warped Hands, Fingers, and Jewelry

Hands remain a classic failure point. Count fingers, watch for digits that merge or split as the hand moves, and check rings, watches, and earrings. Jewelry tends to flicker, change shape, or slide along the skin between frames. Fair warning: frontier models like Sora and Veo have largely fixed static hands, so treat clean hands as inconclusive rather than proof of authenticity.

4. Physics That Do Not Add Up

Generators learn what scenes look like, not how physics works. Watch for objects that float a few pixels above surfaces, limbs that clip through clothing or furniture, hair that ignores gravity, and falls or splashes that play out too slowly or too smoothly. Liquids, smoke, and cloth are especially hard to simulate. If a dropped object lands wrong, your suspicion should rise sharply.

5. Backgrounds That Morph Between Frames

This is the artifact class I see most often in real submissions to our detector. Foreground subjects get the model's attention, while backgrounds quietly warp. Door frames bend, window blinds change count, and a bookshelf rearranges itself over two seconds. Diffusion models struggle with rigid geometry, so straight lines and high-contrast edges behind the subject are the best place to look.

6. Lighting and Shadow Mismatches

Real scenes have one consistent lighting story. Check that shadow direction matches across the subject and background, that skin color temperature fits the environment, and that reflective surfaces show plausible highlights. Face-swap deepfakes are particularly weak here, because a synthetic face gets pasted into real footage and the lighting on the two rarely agrees perfectly.

7. Garbled Text, Logos, and Signage

Generators still write like they are dreaming. Street signs, shop names, jersey numbers, and on-screen captions often come out as letter-shaped gibberish, or as text that changes spelling between frames. Pause on any text in the frame and read it character by character. Real cameras never misspell a stop sign.

8. Audio That Does Not Match the Scene

Close your eyes and just listen. A supposed phone call with zero room echo, no breathing, and no background noise is suspicious because real audio is messy. Synthetic voices also tend to have flat emotional pacing and oddly clean sentence breaks. If the video claims a busy street but the audio sounds like a studio booth, the clip deserves a closer look.

9. Too-Smooth Skin and Repeating Textures

AI skin often looks airbrushed, with pores and fine wrinkles smoothed into plastic. Also scan for repeating patterns: identical leaves on a tree, crowd faces that duplicate, carpet textures that tile visibly. Generators reuse learned texture patches, and repetition is the fingerprint that reuse leaves behind.

The Slow-Motion Test: A 60-Second Manual Workflow

You can run all nine checks in about a minute with this workflow:

1. Save the clip. Download the video or screen-record it so you control playback.
2. Slow it to 0.25x. Most players, including YouTube, support quarter speed. Artifacts that flash past at full speed become obvious.
3. Scrub frame by frame at the cuts. Scene transitions and fast motion are where generators glitch most.
4. Stare at hands, eyes, and edges. These are the three highest-yield zones for artifacts.
5. Screenshot suspect frames. You will want them for reverse image search and for sharing your evidence.

If anything fails these checks, move straight to an automated detector for confirmation.

Check the Video's Source and Metadata

Context catches fakes that pixels hide. Before trusting any clip, check where it came from:

• Reverse search keyframes. Screenshot two or three frames and run them through Google Images or TinEye. If the clip is recycled or staged, earlier versions usually surface.
• Check the uploader. A days-old account with one viral video is a red flag. Established outlets and original sources usually exist for real events.
• Look for content credentials. The C2PA standard attaches signed provenance data showing how media was created. You can inspect a file at the Content Credentials verify tool.
• Check for SynthID. Google watermarks its AI video output and now lets users verify clips in the Gemini app, as described in Google's SynthID verification announcement.

One caveat: social platforms often strip metadata on upload, so a missing credential proves nothing. A present credential, though, tells you a lot.

Why Your Eyes Are No Longer Enough in 2026

Here is the uncomfortable truth from someone who works on detection models daily: perceptual checks are losing. Research backs this up. A study published in iScience by Köbis and colleagues, titled "Fooled twice: People cannot detect deepfakes but think they can," found that participants could not reliably distinguish deepfakes from authentic video, and overestimated their own ability to do so (Köbis et al., iScience, 202101335-3)).

Against frontier models, the nine signs above shrink to a handful of subtle tells. Manual checks are your fast first pass, not your final answer. For anything that matters, money, reputation, or news, you need automated analysis.

Verify Any Video with an AI Video Detector

Automated detectors see what eyes cannot. Our AI video detector analyzes a clip frame by frame, checking temporal consistency between frames, hunting for generation fingerprints in the signal, and matching artifacts against known models like Sora and Runway. It reaches high accuracy across our internal benchmarks.

Using it takes three steps:

1. Upload the clip or paste a link to it.
2. Wait a few seconds while the analysis runs across multiple detection models.
3. Read the verdict (Authentic, Likely Synthetic, or Inconclusive) with its TrustScore from 0 to 100.

The free tier includes 50 detections per month for clips up to 2 minutes long per detection (paid plans handle up to 10 minutes), and files are deleted from primary storage within 60 seconds of analysis completion, so your uploads are not stored or reused unless you opt into retention. If you want to compare options first, see our roundup of the best AI video detectors, or start with our free deepfake detector.

Not sure about a video? Scan it free: 50 checks per month, files deleted after analysis. Try the AI video detector

FAQ

Can you tell if a video is AI generated for free?

Yes. The manual checks in this guide cost nothing, and our detector's free tier includes 50 detections per month. That combination covers almost any consumer use case without spending a dollar.

What is the most reliable sign a video is AI generated?

No single sign is reliable on its own. The strongest combination is lip-sync drift plus physics errors plus a detector's confidence score. Treat each sign as evidence, not as a verdict.

Can AI detectors be wrong?

Yes. Detectors return confidence scores, not certainties, and false positives happen, especially with heavily compressed or edited footage. That is why a good result includes a verdict and a confidence score rather than a bare yes or no.

Does TikTok or YouTube label AI videos?

Partially. Both platforms require creators to disclose realistic AI content and apply labels based on self-disclosure and watermark signals. Enforcement is inconsistent, so an unlabeled video is not evidence that it is real.

Can you detect Sora videos?

Yes. Detectors target the fingerprints each generation model leaves in its output, and Sora-class video carries detectable patterns. It is an arms race, though: detection models need constant retraining as generators improve.

Conclusion: Trust but Verify Every Video

Knowing how to tell if a video is AI generated comes down to a two-stage habit. First, run the manual pass: slow the clip down, check the nine signs, and trace the source. Second, verify anything important with an automated tool, because 2026-era generators routinely beat human eyes.

The same layered approach applies to face swaps, which we cover in our guide on how to detect deepfake videos. And if you are still getting oriented, start with what a deepfake video is.

Saw something suspicious today? Run it through the AI video detector now. 50 free checks per month, no card required.

Related reading