How to Tell If a Video Is AI-Generated (Free Online Tool)

Why Your Instincts Are No Longer Enough

Human pattern recognition is remarkably good — but it was built for a different era. We evolved to detect fakes in a world where creating a convincing forgery required real skill and time. The mental shortcuts we use — does the face move naturally? does the voice match the lips? does the background look real? — were reliable for decades.

AI has systematically eliminated every one of those proxies.

Models like Sora, Runway, and Kling are trained on millions of hours of real footage. They don't just copy human movement — they statistically model it, producing output that satisfies the same perceptual expectations our brains use to judge authenticity. The result is footage that passes casual scrutiny, reverse image search, and in many cases, professional review.

Research confirms what most people are reluctant to admit: human accuracy at detecting AI-generated video hovers around 50–60% under controlled conditions. For the most realistic outputs, that's barely better than a coin flip. And in real-world conditions — quick scroll, low attention, emotional context — performance drops further.

This isn't a reason to panic. It's a reason to stop relying on instinct alone and use tools built for the job.

What AI Video Detection Actually Checks

Manual detection looks for visible artifacts — things a human eye can spot. Automated detection goes several layers deeper, analyzing properties of the video that are invisible to casual viewing but statistically distinctive.

Temporal consistency analysis

Real cameras capture motion with specific physical properties: motion blur, rolling shutter, grain patterns tied to sensor behavior. AI generators approximate these properties but don't replicate them perfectly. Frame-by-frame inconsistencies that are invisible at normal playback speed become detectable when analyzed statistically across the full clip.

Frequency domain fingerprinting

Every AI generation pipeline leaves characteristic signatures in the mathematical structure of the image data — patterns introduced by the model's architecture that persist even after compression and re-encoding. These signatures aren't visible. They exist in the frequency spectrum of the video and require model-based analysis to surface.

Facial geometry and landmark tracking

Even when individual frames look photorealistic, the trajectory of facial landmarks — the precise spatial relationships between eyes, nose, mouth, and jaw — over time often differs subtly from real human movement. Real faces move with muscle-driven physics. AI faces are generated frame by frame, and the continuity is approximated rather than physically grounded.

Metadata and provenance analysis

AI-generated videos frequently carry metadata that doesn't match their claimed origin. Missing camera model fields, inconsistent creation timestamps, encoder strings characteristic of specific generation pipelines — none of these are things a viewer would notice, but they're readable by software in milliseconds.

UncovAI's video detection runs all of these checks simultaneously. The result is a confidence score based on converging signals — not a single heuristic that can be gamed by adjusting one parameter.

The Situations Where Getting This Wrong Is Costly

Most people encounter AI-generated video in low-stakes contexts first — a viral clip, a meme, a social post — and assume their ability to spot fakes there translates to higher-stakes situations. It doesn't. The same models that produce obviously artificial content can produce deeply convincing forgeries when the goal is deception rather than entertainment.

How to Check If a Video Is AI-Generated

Option 1: Use UncovAI (recommended)

Paste a link or upload a file at UncovAI's video detection tool. The scan runs in under ten seconds and returns a verdict with a confidence score. No account required. Works on videos from Instagram, TikTok, YouTube, WhatsApp exports, and direct file uploads. If you want to scan on the go, the UncovAI browser extension lets you check any video without leaving the page you're on.

Option 2: Inspect the metadata manually

For MP4 and MOV files, metadata can be read using free tools like ExifTool or MediaInfo. Look for blank camera make and model fields, creation timestamps that don't match the claimed date, and encoder strings associated with known generation tools. This is a useful secondary check, but metadata can be stripped or spoofed — it's not a substitute for model-based detection.

Option 3: Step through the video frame by frame

Download the video and use VLC — press E to advance one frame at a time. Focus on three areas: teeth geometry (AI models frequently produce frames where teeth merge, multiply, or shift shape), background object behavior (items that subtly change position between frames), and hair edges (real video shows consistent motion blur at the hairline; AI video often produces frame-to-frame inconsistency at fine edges). This is slow, requires practice, and misses the frequency-domain signatures that software catches. Use it as a last resort or a sanity check.

Option 4: Reverse video search

Google Lens and TinEye can sometimes identify the original source of re-edited content. If a video claimed to be original turns out to be a real clip manipulated with AI, a reverse search may surface the source. Limitation: this only catches re-used content. It finds nothing for original AI generation, which is increasingly the norm in deliberate deception.

For anything involving AI-generated audio or voice cloning alongside video, run a separate audio scan — the two signals together provide much stronger evidence than either alone.

Frequently Asked Questions