Sora, Veo 3, and Kling: How to Detect AI-Generated Video in 2026

The Video Threshold Has Been Crossed

For most of AI's history, video was the frontier that would not fall. Images were convincing by 2022. Audio was convincing by 2023. But video, with its requirement for temporal consistency across thousands of frames, realistic physics simulation, and coherent lighting across moving scenes, was considered the last reliable signal of authenticity.

That changed in 2024. OpenAI's Sora demonstrated that a text prompt could produce 60-second video clips with cinematic quality. By 2025, Google's Veo 2 and the Chinese model Kling had joined the field, each producing video that major news organizations initially published without verification. In 2026, Veo 3 can generate video with synchronized audio (footsteps, ambient sound, dialogue), all synthesized from a single text description.

The era of trusting your eyes for video has ended. But the forensic tools are keeping pace.

Why Video Deepfakes Are Harder to Detect

A static image analysis looks for spatial inconsistencies: lighting from the wrong angle, noise patterns that don't match hardware signatures, compression artifacts that reveal manipulation. Video introduces a time dimension that multiplies both the detection challenge and the number of available signals.

AI video models are trained to produce frames that are individually plausible. The failure modes appear in the transitions between frames-in the physics of how objects move, in the consistency of lighting as a camera pans, and in the subtle flickering of fine details like hair, fabric texture, and background elements.

The Key Forensic Tells of AI Video

1. Temporal Flickering

Look at a single small region of a suspected AI video, a patch of wall, a section of sky, a square centimeter of clothing texture. In a real video, this region will be stable between frames (allowing for camera shake and motion blur). In an AI-generated video, static regions often exhibit subtle frame-to-frame variation, a barely perceptible flickering that is the model regenerating each frame independently rather than tracking continuity from the previous frame.

This is most visible on solid surfaces. Pause the video on consecutive frames and compare. Real footage is consistent. AI footage often 'breathes.'

2. Physics and Fluid Dynamics

Current AI video models perform poorly at simulating the behavior of fluids, fabric, and particulates. Water does not flow with the correct viscosity. Fabric does not follow realistic drape physics under gravity. Smoke and dust expand according to aesthetic rules rather than thermodynamic ones.

This is perhaps the most reliable tell because it is extremely difficult to fake. Physics simulation requires solving differential equations that neural networks currently approximate poorly. A video of someone walking through rain, for example, will often show rain that bounces at incorrect angles from surfaces, droplets that are uniform in size (real rain has a wide size distribution), and puddle ripples with the wrong radius-to-frequency ratio.

3. The 'Boiling' Background Effect

AI video models allocate most of their computational attention to the primary subject, the person or object specified in the prompt. Background elements are rendered with lower fidelity and often exhibit what researchers call 'boiling': a subtle, chaotic shifting of texture and detail that makes background foliage, crowds, or architectural detail appear to vibrate slightly.

4. Compression Artifact Patterns

When an AI-generated video is encoded for distribution (typically with H.264 or H.265 compression), the codec interacts with AI-generated content in distinctive ways. The blocky artifacts that appear in highly-compressed real video follow the structure of the underlying image. In AI video, these blocks often cluster at AI-generated boundary regions, the edges between synthesized objects, creating a characteristic artifact pattern different from natural compression.

5. Facial Microexpressions and Blink Patterns

Early face-swap deepfakes rarely blinked, a tell that trained observers quickly learned to look for. Modern systems blink convincingly. But they still fail on microexpressions: the sub-100-millisecond muscle contractions around the eyes, mouth, and forehead that accompany speech and emotion in real humans. Forensic video analysis can measure the timing and symmetry of these microexpressions. AI faces are often too symmetrical-real faces are subtly asymmetric, and that asymmetry is dynamic.

The Arms Race: Where Video Detection Is Heading

Video detection is an active research field, and the honest summary of its current state is less flattering than most vendor pages suggest. The Deepfake-Eval-2024 study measured detector performance on real-world media rather than curated academic sets and found that open-source video detectors lost roughly half their measured AUC in the transition, concluding that academic benchmarks are out of date and not representative of real-world deepfakes. Any accuracy figure quoted for AI video detection almost certainly comes from the first kind of test rather than the second.

The key emerging approach is model-agnostic rather than model-specific, which matters because the generator landscape turns over faster than any fingerprint library can track.

The key emerging approach is temporal consistency network analysis: rather than analyzing individual frames for spatial artifacts, these systems analyze entire clips for physics and continuity violations across time. This approach is model-agnostic; it does not need to know which AI system generated the video, only that the video violates the laws of physical reality in measurable ways.

What a Re-Encode Does to the Evidence

This is the part almost nobody writes about, and it is the most common reason a video verdict goes wrong in practice.

Almost no video reaches you in the state its generator produced it. It gets saved to a phone gallery, sent through a messaging app, re-shared, screen-recorded, or run through a platform transcoder. Every one of those steps re-encodes the file, and re-encoding does two things at once: it degrades the pixel-level evidence a detector depends on, and it overwrites the container metadata with the signature of the device that did the re-encoding.

The second effect is the dangerous one, because it does not look like degradation. It looks like provenance. An AI-generated clip that has been saved to an Android gallery and shared onward will carry a device-native frame rate, a mobile handler string, a mono audio track and a variable frame rate profile - the exact combination a naive check reads as evidence of a real camera. The file has not become more authentic. It has acquired a costume.

We learned this the hard way. An early version of our own video pipeline scored container metadata as though it were forensic evidence, and it could return a confident authenticity verdict on a re-shared AI clip without ever extracting a frame. One reporter had seven AI videos certified as genuine. The fix was not to retune a threshold - it was to stop letting the container vote on a question it cannot answer.

The lesson generalises well beyond our own tooling: if a video verification tool tells you a clip is authentic and its reasoning refers to camera or device metadata, it has told you about the last program that touched the file, not about who made the footage. Ask what the pixels say.

How a Multi-Signal Video Analysis Actually Works

Single-frame analysis and video analysis are different problems, and a tool that simply runs an image detector on a thumbnail is doing the first while claiming the second.

A defensible video pipeline works in two passes. The first extracts keyframes across the clip and runs the full forensic stack on each one independently - compression history, sensor noise, frequency-domain artefacts, lighting consistency. That produces a set of per-frame readings rather than a single score.

The second pass is the one that only video makes possible: comparing those frames against one another. Sora, Veo, Kling and Runway Gen-3 all produce frames that are individually plausible and that drift slightly relative to each other over time. A face that is subtly the wrong width in frame 90 relative to frame 30. A background sign whose letters reorganise. A shadow whose angle wanders without the light source moving. None of these is visible in any single frame, which is exactly why a still-image detector cannot see them.

The two passes disagree usefully. When per-frame signals are clean but inter-frame drift is high, you are usually looking at a well-rendered generated clip. When per-frame signals fire but drift is low, you are more often looking at real video that has been edited, upscaled or partially repainted - a different finding with different consequences, and one a single blended score would hide.

Where Video Detection Fails

Being specific about failure conditions is more useful than a headline accuracy number, so here are the ones that matter.

Short clips. Inter-frame analysis needs frames. A three-second clip gives far less temporal evidence than a thirty-second one, and below a couple of seconds the second pass has almost nothing to work with.

Heavy compression. A clip that has been through several platform transcodes has had much of its pixel-level evidence destroyed. Frequency-domain artefacts survive re-compression better than compression-history analysis does, but nothing survives indefinitely.

Screen recordings. Recording a video off a screen resamples every frame and inserts the recording device's own characteristics. This defeats container-level checks entirely and degrades pixel-level ones substantially.

Short-form platform output. Clips that have passed through the aggressive re-encoding and stabilisation pipelines of short-form video platforms arrive with their evidence degraded and their metadata rewritten - the two failure modes above, stacked.

Hybrid footage. Real video with an AI-generated element composited in, or AI footage with real audio, sits between categories. A binary verdict is the wrong output shape for it, which is why per-signal evidence matters more than a single label.

Frequently Asked Questions

Can I tell if a video is AI-generated just by looking at it?

Increasingly, no. Human accuracy on high-quality synthetic media sits close to chance, and the tells trained observers learned for early deepfakes - absent blinking, obvious warping at face boundaries - have largely been engineered away. The reliable tells that remain, such as fluid dynamics and inter-frame drift, are difficult to assess by eye in real time.

Does a video with real camera metadata prove it is real?

No, and this is the most common mistake. Re-encoding a clip - by saving it to a phone gallery, sharing it through a messaging app, or letting a platform transcode it - stamps that file with the signature of the device that did the re-encoding. Metadata describes the last program to touch the file, not the origin of the footage.

Why do detectors disagree with each other?

Because they measure different things. Some analyse a single extracted frame, some analyse compression history, some look for fingerprints of specific named generators, and some analyse temporal consistency. A tool trained on the fingerprints of one generation of models will underperform on a generator released after its training data closed. Disagreement between tools is informative rather than embarrassing, and it is a good reason to read the per-signal evidence rather than only the verdict.

What file should I submit for the most reliable result?

The original file, as close to source as you can get it, and the longer the clip the better. Every re-share degrades the evidence. If you have a choice between a forwarded copy and the original upload, always use the original - and if you only have a screen recording, treat any result as substantially less reliable.

Does a no-AI-detected result mean the video is genuine?

It means the analysis found no reliable evidence of AI generation at the time of scanning, which is a narrower statement. A heavily compressed clip, a very short clip, or footage from a generator unlike anything in the fingerprint library can all return that result. Treat it as an absence of evidence rather than evidence of absence, and weigh it against whatever else you know about where the video came from.

What You Should Do Right Now

If you encounter a dramatic or emotionally charged video on social media, particularly one that appears to show a crisis, a political figure, or a celebrity in an unusual situation-apply skepticism proportional to the stakes. Reverse image search individual frames using Google Lens. Check whether the audio is synchronized with lip movements at a frame-by-frame level. Look for the tells described above. And run suspicious clips through algorithmic analysis before sharing or reporting them.

The burden of verification has been democratized. Professional forensic tools are no longer confined to research labs. An AI video detector is now available to anyone willing to spend 60 seconds before hitting the share button.