How AI video detection works

A video gives you two kinds of evidence, and the second one is the useful one.

The first is what you would get from any still. We extract key frames and run the full six-signal analysis on each: GAN fingerprint detection, Error Level Analysis, PRNU sensor-noise fingerprinting, frequency-domain analysis, metadata forensics and neural classification. A frame from a generated video is, on its own, a generated image, and it carries the same artifacts.

The second is temporal, and it has no equivalent in photo forensics. Video generators produce frames that are individually convincing and collectively a little inconsistent, because coherence across time is something the model approximates rather than something physics enforces. Real footage gets consistency for free: the same camera, the same sensor, the same scene, moving continuously. A generated clip has to reconstruct that agreement on every frame, and it never quite manages.

So we compare frames against each other and look for the failures. Edges that shift slightly between frames when nothing in the scene moved. Texture that reorganises rather than translates. Lighting that changes direction by a degree or two across a second of footage. Facial landmarks that jitter when the head is still.

Face-swapped video fails in a different and more obvious way. The underlying footage is genuine, so the background is temporally perfect while the swapped region is not. That mismatch between a stable scene and an unstable face is one of the strongest signals available, because it requires no reference and no comparison to anything outside the file.

The result comes back per signal and per analysed segment, so a clip that is real for twenty seconds and synthetic for three shows up that way instead of being averaged into a single unhelpful number.

What each generator gives away

Each generator has a different architecture, and architecture leaves a signature.

Sora, from OpenAI, is a video diffusion transformer. It produces strong spatial quality with characteristic temporal coherence artifacts at object boundaries, and its output container has historically carried metadata that identifies it outright.

Veo, from Google DeepMind, was trained at YouTube scale and shows it. Motion synthesis is smoother than most competitors, which paradoxically helps: real handheld footage has micro-jitter that Veo tends to under-produce, so the motion is too clean.

Kling, from Kuaishou, has a different training distribution and gives itself away in backgrounds. Crowds, foliage and architecture behind the subject reorganise between frames in ways a real scene does not.

Runway Gen-3 is strongest on short, tightly framed shots and weakest as clips lengthen, with drift accumulating in anything held on screen for more than a few seconds.

Two honest caveats. New model versions ship faster than any detector updates, so a brand-new generator will always have a window where accuracy on its output is lower than on established ones. And attribution is a weaker claim than detection: telling you that a clip is synthetic is a more reliable statement than telling you which specific model made it. The evidence chain distinguishes the two rather than presenting both with the same confidence.

Why time is the giveaway

A single AI-generated image carries generation artifacts in its pixels. A video adds an entirely new attack surface: the relationship between frames over time.

Current AI video generators struggle with object permanence - keeping every visual element consistent from one frame to the next. A person's shirt may shift color slightly between frames. A logo in the background may subtly change shape. Hair strands behave as if guided by different physics rules at different moments in the clip.

Physics consistency is another reliable signal. AI video generators simulate fluid dynamics, cloth movement, and hair response using learned approximations rather than physical models. Water splashes follow trajectories that are statistically plausible but not physically precise. Cloth folds in wind settle too quickly or too symmetrically. These deviations are measurable.

Motion blur is a third temporal signal. Real camera footage captures motion blur that is directionally consistent with the actual direction of movement. AI-generated motion blur is frequently rendered with incorrect directionality, blur that does not match the motion vector of the subject, a pattern detectable through optical flow analysis.

Eye blink timing in deepfake faces is measurably abnormal. Real people blink roughly 15 to 20 times per minute, with each blink lasting 150 to 400 milliseconds. Deepfake faces frequently skip blinks entirely or insert them at algorithmically distributed rather than neurologically natural intervals. FauxLens measures these temporal signals alongside per-frame forensics, which is why our video detection catches content that per-frame analysis alone would miss.

What this is being used for

AI video has moved from research curiosity to active fraud tool. The most documented case is the 2024 Hong Kong CEO fraud incident, in which criminals used AI-generated video of senior executives in a live video call to convince a finance employee to transfer HK$200 million - equivalent to $25.6 million USD. The employee believed the faces and voices were real because the call appeared indistinguishable from a legitimate meeting. This was the first large-scale, documented financial fraud in which AI-generated video of specific individuals was used in real time to authorize a wire transfer.

In political disinformation, AI-generated video of politicians making statements they never made circulated on social media ahead of multiple national elections in 2024 and 2025. These clips were produced using face-swap technology and voice cloning, then distributed through coordinated networks before fact-checkers could respond. Several clips reached millions of views before removal.

Journalists working in conflict zones now encounter AI-generated evidence footage submitted by multiple parties - videos purporting to show atrocities, troop movements, or civilian casualties that cannot be verified through on-the-ground reporting alone. Veo 2-generated synthetic footage has been identified in documented disinformation operations in 2025.

Security teams at financial institutions now screen video calls for deepfake filters as standard fraud-prevention procedure. Legal professionals are beginning to submit AI video analysis reports as part of evidence authentication workflows. Journalists, lawyers, and security teams all need reliable video verification - which is exactly what FauxLens provides.

Getting a usable result

The quality of your upload directly affects detection accuracy. Follow these guidelines for the most reliable results.

Trim to the suspicious section. FauxLens analyzes the full clip you upload. If you have a 10-minute interview and you suspect the first 2 minutes, upload those 2 minutes rather than the full recording. This reduces processing time and focuses the forensic analysis on the content that matters.

Upload at original quality whenever possible. Video that has been re-encoded, compressed for social media sharing, or converted between formats loses information that our forensic engine uses. The original file from a recording app or a direct platform download provides the strongest signal.

MP4 versus MOV: both formats are fully supported. MP4 is the standard for most device recordings. MOV files from Apple devices are accepted without conversion. Do not convert between formats before uploading - conversion introduces re-encoding that degrades forensic signals.

Bitrate matters more than frame rate for detection quality. A 30fps clip at a high bitrate is more analyzable than a 60fps clip that has been aggressively compressed. Prioritize bitrate over frame rate when you have a choice.

If your file exceeds the 100MB upload limit, trim the clip further using a lossless cutting tool that does not re-encode the video, such as FFmpeg with the -c copy flag. Re-encoding to reduce file size is acceptable as a last resort but will reduce some detection signal strength.