How to Tell If a Video Is AI-Generated: A Detection Guide

Why AI video verification has become a routine skill

A few years ago, spotting a synthetic clip was easy. Faces melted, hands multiplied, and backgrounds shimmered like wet paint. That era is over. Modern generative video models produce coherent motion, believable lighting, and lip-sync that holds up on a phone screen. The result is a simple but uncomfortable truth: casual viewing is no longer a reliable verification method.

At the same time, the volume of synthetic video has exploded. Marketing teams use generated footage for ads and product explainers. Newsrooms receive clips from anonymous accounts. HR and legal teams review recorded interviews. Social platforms process millions of uploads per day with limited human review. In each of these contexts, someone eventually has to answer a specific question: was this recorded from reality, or was it generated, edited, or manipulated by a model?

That question rarely has a clean yes-or-no answer. Verification is almost always a probability judgment built from multiple weak signals. A missing watermark means little on its own, because a single re-encode erases most embedded markers. A detector that returns "92% AI" is not a verdict; it is one data point from a system that has never seen the exact model that produced the clip.

This guide lays out a practical, tool-agnostic workflow for checking whether a video is AI-generated. It covers the technical traces models leave behind, the manual inspection steps that catch what automation misses, the limits of detection software, and how to turn all of it into a repeatable process for teams that handle video at scale.

How AI video generators leave traces behind

Every generated clip carries evidence, but the evidence lives at different layers: the file container, the pixels, the motion over time, and the audio track. A thorough review touches all four.

Provenance metadata and invisible watermarks

Some generators and distribution platforms attach provenance records to a file. The most widely adopted approach is content credentials based on the C2PA standard, which stores a signed manifest describing how a piece of media was captured or created, and what edits followed. Similar systems embed invisible watermarks into pixel data or audio, designed to survive mild compression.

These records are the strongest evidence you can find, because they are declarative rather than inferential. But they are also fragile. Uploading to a social platform, converting formats, cropping, screen-recording, or re-encoding for delivery typically strips metadata and damages embedded watermarks. So the practical rule is: check provenance first, and treat its absence as neutral, not as proof of authenticity.

Pixel-level and statistical anomalies

Generators approximate reality rather than simulate it. That approximation leaves statistical fingerprints. Common visual tells include:

Noise inconsistency. Real sensors produce a fairly uniform grain pattern across a frame. Generated footage often shows noise that changes between regions, or disappears entirely in flat areas like walls and skies.
Texture smearing. Fabric weaves, foliage, brickwork, and hair strands blur into plausible-looking mush when you zoom in past 200%.
Edge blending artifacts. Where a subject meets the background, generated clips frequently show a soft halo, a thin dark line, or a subtle vibrating edge.
Text and logos. Background signage, screen text, and product labels often contain slightly wrong letterforms or characters that shift between frames.
Optic inconsistency. Lens flare, depth of field, chromatic aberration, and motion blur that do not match a single real camera setup — or that change mid-shot.

Motion, physics, and object behavior

Temporal analysis catches things a single frame hides. Watch for objects that morph shape rather than move, limbs that pass through clothing or furniture, and reflections in mirrors or windows that do not match the subject. Crowds are a classic weak point: background figures may glide rather than walk, or blend into each other on contact.

Camera behavior matters too. Real handheld footage has micro-jitter tied to breathing and footsteps. Generated camera moves are frequently too smooth, or they drift in ways a physical operator could not replicate — a slow push-in that never triggers parallax between foreground and background, for example.

Audio and lip-sync tells

If the clip has sound, listen closely with headphones and inspect a spectrogram. Synthesized speech tends to have flattened dynamics: consistent volume, thin room tone, and unnatural gaps between words. Real recordings carry breath, mouth noise, and environmental consistency that changes as the speaker moves relative to the microphone. Lip-sync that is fractionally early or late, or phonemes that do not quite match the mouth shapes, are strong indicators — especially in side-profile shots where generators struggle.

A step-by-step manual verification workflow

You do not need a forensic lab to run a credible check. You need a consistent sequence so you do not skip an obvious signal.

Step 1: Secure the original artifact

Before anything else, preserve the file exactly as received. Do not re-export, do not trim, do not upload it anywhere yet. Copy it to working storage, record the file hash, and note where it came from, when it arrived, and who sent it. Every subsequent inspection should run against that untouched original, because re-saving can destroy metadata and watermark evidence you will never get back.

Step 2: Run a metadata and provenance pass

Inspect the container with a metadata tool that shows streams, codecs, encoder strings, creation timestamps, GPS data, and embedded manifests. Look for three things: whether a signed provenance manifest exists, whether the encoder string names a known generation pipeline, and whether the technical properties make sense. A clip claiming to be from a phone camera but carrying a resolution, bitrate, and codec combination typical of web export is worth questioning.

Step 3: Inspect frames at full magnification

Extract a spread of frames — roughly every half second for short clips — and examine them at 100% and 400% zoom. Check hands, ears, teeth, jewelry, glasses, and any reflective surface. Look at background text and repeating patterns. Pay attention to where the subject touches the environment: feet on ground, hands on tables, shadows under chairs. Generated contact points often float slightly or blend unnaturally.

Step 4: Watch motion in slow motion

Play the clip at 25% to 50% speed and step frame by frame through any movement. Count limbs. Watch how fabric and hair react to motion. Check whether shadows move consistently with their objects. Look for flicker, warping, or sudden changes in identity between adjacent frames — a telltale sign of a short generated segment stitched into longer footage.

Step 5: Analyze the audio track

Isolate the audio and check it in three ways: listen with headphones for breath and room consistency, view a spectrogram for cut-off frequencies and unnatural regularity, and compare lip movements against the waveform. If the audio was generated separately, you will often see a slight mismatch between plosive sounds and visible mouth closures. If it was cloned from a real speaker, the emotional range is frequently narrower than the subject's other recordings.

Step 6: Corroborate externally

Search for the clip, or keyframes from it, across the web. Locate the original upload and check whether the account has a history. Look for the same scene from a second angle or a second source. Verify whether the location, weather, event, and clothing make sense for the claimed date. Corroboration is often the fastest route to a confident conclusion, because it bypasses the technical arms race entirely.

Step 7: Score, document, and decide

Write down each signal you found, what it suggests, and how strong it is. Then apply your rubric (below) to reach a public-facing conclusion. Documenting the process matters as much as the conclusion: if your call is challenged later, the record of what you checked and when is your defense.

A scoring rubric for consistent decisions

Binary verdicts create unnecessary conflict. A weighted score keeps reviewers aligned and makes uncertainty explicit. Below is a starting rubric you can adapt. Assign 0–3 points per category, where 0 means no concern and 3 means strong evidence of synthesis.

Signal category	What you check	Weight
Provenance	Signed manifest, watermark, encoder strings	High
Visual anatomy	Hands, teeth, ears, reflections, text	High
Temporal motion	Morphing, physics errors, identity flicker	High
Audio	Room tone, breath, phoneme-lip alignment	Medium
Contextual plausibility	Location, weather, event, account history	Medium
Detector output	Multiple independent detectors agreeing	Low to medium

Interpretation guidance: a high-weighted signal plus two medium signals usually justifies a public statement that the clip is likely synthetic. A single medium signal plus a detector score does not. When the total lands in the middle, say so explicitly — "we could not verify the origin of this clip" is a legitimate and often correct conclusion.

Tools and standards worth knowing

You do not need a single magic detector. A small toolkit covers almost everything.

Container and metadata inspectors. Command-line metadata utilities and media info tools reveal codecs, bitrates, encoder strings, timestamps, and embedded manifests.
Frame extraction and playback. A standard video toolkit lets you extract frames at precise intervals and step frame by frame, which is essential for temporal analysis.
Audio analysis. A free waveform and spectrogram editor is enough to spot frequency cutoffs, unnatural silence, and compression patterns.
Provenance standards. Content credentials and similar signed-manifest systems let you verify a chain of custody when the file still carries it.
Detector services. These are useful as corroboration, not as verdicts. Run at least two from different vendors and note when they disagree — disagreement is itself informative.
Reverse search and keyframe lookup. Searching individual frames rather than whole files dramatically improves the odds of finding the original source.

Where detection fails: limits, false positives, and adversarial edits

Detection has a hard ceiling, and understanding it prevents embarrassing mistakes.

Re-encoding destroys evidence. A clip downloaded from a messaging app has usually been transcoded, resized, and stripped of metadata. Watermarks may be gone. You are then working with pixels and motion only, which raises the error rate substantially.

Detectors are probabilistic and biased. Most detector models are trained on outputs from a limited set of generators. Newer or less common models slip past them. Meanwhile, heavily color-graded real footage, slow-motion phone video, denoised archival film, and stylized animation are routinely flagged as synthetic. A detector output is a hint about model familiarity, not a property of the video.

Real footage can be manipulated without being generated. Cutting, speed ramping, face swapping onto a real performance, or adding a generated background all produce hybrid clips. These break the binary framing of "real versus AI" and require you to describe what specifically was altered.

Adversarial editing is cheap. Slight noise injection, resizing, or a single compression pass can defeat many watermark schemes. Assume that anyone motivated to hide synthesis can do so at low cost.

Common mistakes during video verification

Uploading the file before inspecting it. Every platform upload is a chance to lose provenance data permanently.
Judging from one frame. A single suspicious hand proves nothing. Patterns across frames matter.
Treating a detector percentage as truth. A number without methodology is not evidence.
Ignoring audio because the visuals look fine. Cloned voice is often the easiest signal to confirm.
Skipping context. A clip that contradicts the weather, the location, or the calendar is questionable regardless of pixel quality.
Not recording your process. Undocumented conclusions cannot be defended or repeated.
Confusing synthetic with false. A generated clip can illustrate something true, and a real clip can be staged. Separate the production method from the claim being made.
Announcing certainty you do not have. Overclaiming in either direction damages trust more than admitting uncertainty.

Turning detection into a repeatable team process

For individuals, a checklist is enough. For organizations publishing or moderating video, the checklist needs to become a workflow with owners.

A workable structure looks like this. Intake is handled by whoever receives the file; they preserve the original, log the source, and run the metadata pass. A reviewer performs the visual, temporal, and audio inspection and records findings against the rubric. A second reviewer independently reviews anything scoring above a defined threshold. A designated decision-maker — an editor, trust and safety lead, or legal contact — makes the publish, label, or reject call. Everything is stored with the original file and the hash so the audit trail survives.

Two policies make this sustainable. First, define your disclosure standard in advance: when do you label content as containing synthetic elements, and how do you phrase it? Second, define an escalation path for legal or safety-sensitive material, so reviewers are not making judgment calls alone under pressure.

FAQ

Can any tool give a definitive answer about whether a video is AI-generated?
No. Every available method produces a probability estimate. The closest thing to certainty is cryptographically signed provenance data attached at creation and preserved through the entire chain of distribution — which is rare in practice.

How accurate are AI video detectors?
Accuracy varies enormously by generator, compression level, and content type. Detectors tend to perform well on outputs from the models they were trained on and poorly on everything else. Treat them as supporting signals.

What is the fastest manual check?
Zoom into hands, teeth, ears, and any reflective surface, then watch the clip at quarter speed. These two steps catch a large share of synthetic footage in under a minute.

Does removing metadata prove a video is fake?
No. Messaging apps, social platforms, and video editors strip metadata routinely. Missing provenance is weak evidence, not proof.

Why do generated videos often look fine on a phone but strange on a monitor?
Small screens hide fine texture errors, edge blending artifacts, and motion inconsistencies. Reviewing on a larger display at full resolution materially improves detection.

Should I label content that merely uses AI for editing?
That depends on your disclosure policy. A useful rule is to disclose whenever synthetic elements could mislead a viewer about what was physically recorded.

What should I do if I cannot reach a conclusion?
Say that. Describe what you checked, what you found, and why the evidence was inconclusive. A documented "unverified" is more valuable than a confident guess.

Key takeaways

Verification is a layered process, not a single test. Provenance, pixels, motion, audio, and context each contribute partial evidence.
Check metadata and provenance before you touch, convert, or upload the file. That window closes quickly.
Slow-motion review and high-zoom frame inspection remain the highest-value manual techniques.
Detector scores are corroboration, not verdicts, and disagreement between detectors is meaningful information.
Use a weighted rubric so conclusions are consistent, defensible, and honest about uncertainty.
Build the process into your team's workflow: preserved originals, documented findings, an independent second look, and a clear policy for labeling synthetic media.