How to Detect an AI-Generated Video in 2026 (Sora, Kling, Runway)

A video of a public figure says something they never said. A job applicant’s interview call doesn’t quite match the photo on their CV. A viral clip of a disaster looks slightly wrong in ways you can’t name. In 2026, the question “is this video AI-generated?” comes up daily — and the honest answer is that your eyes in real-time playback are no longer sufficient to settle it.

Text-to-video models (Sora 2, Kling 2, Runway Gen-4, Google Veo) have reached a quality level where a casual watch will not catch them. Face-swap deepfakes have been deployable by non-experts for several years. The visual tells still exist — but they require you to slow down, go frame by frame, and know what to look for. This guide covers both families of AI video, the signals that distinguish them, and how to verify a suspicion systematically.

Text-to-video vs. face-swap: two different problems

Before reviewing specific signals, it helps to understand the two families of AI video, because their artifacts differ.

Text-to-video generators (Sora 2, Kling, Runway Gen-4, Google Veo) synthesize every pixel from scratch based on a prompt. They are improving rapidly and now produce clips that appear coherent at normal viewing speed. Their artifacts tend to live in physics, temporal consistency, and fine detail — things that are hard to synthesize correctly but easy for the human visual system to accept if you’re not looking carefully.

Face-swap deepfakes graft a target face onto existing footage using a trained face-swap model. The original scene, lighting, and motion are real; only the face region has been replaced. This creates a different artifact profile: the boundaries around the face carry compression and blending seams, the skin texture of the swapped region may not match the original, and lip movement may not align with the audio.

Both produce forged content. Both leave forensic traces. The detection approach overlaps but is not identical.

1. Watch for temporal flicker and inconsistency

The most distinctive class of AI video artifact is temporal: something that changes or jumps between frames in a way that physical reality does not allow.

Text-to-video models generate each frame semi-independently. Even with optical flow conditioning, fine details — a strand of hair, the edge of a collar, a reflection in a window — can shift position or change between adjacent frames in ways that no natural motion would produce. At 24 or 30 fps this reads as a subtle shimmer or flicker on stationary objects.

To find this: pause the video and step through it one frame at a time. Look for regions where a stable background element or stationary detail changes between frames. Pay particular attention to:

• Texture in fabric, skin, and hair against a still background
• Edges where the subject meets the scene
• Reflections in glasses, mirrors, or water
• Any text visible in the frame

2. Check physics and motion

Generative models learn statistical patterns of motion, not physical laws. This produces specific failures:

• Cloth and hair dynamics — fabric folds that don’t respond to the motion of the body wearing it; hair that moves in the wrong direction relative to head movement or apparent wind.
• Rigid body physics — objects that pass through each other or change mass mid-clip (a bag handle that cuts through an arm, a drink that doesn’t slosh when the table moves).
• Fluid and smoke — water, smoke, and fire are among the hardest dynamics to synthesize. Fluids in AI video tend to look looped, symmetric, or unnaturally smooth.
• Gravity — objects falling at the wrong rate, or debris from a fall that curves when it should drop straight.

These failures are easy to miss at normal playback speed. Scrubbing slowly or using a frame-advance control is the method here.

3. Inspect face boundaries and skin texture

This is the primary signal in face-swap deepfakes and remains relevant in text-to-video content that includes people.

For face-swap content, the replaced face region has been composited into the original footage. Look at:

• The boundary between the face and the neck, ears, and hairline — blending artifacts appear here as a slight halo, uneven skin tone, or unnaturally smooth transition.
• Skin texture consistency — the swapped face region may be slightly blurrier or smoother than the rest of the frame, particularly in high-detail areas like pores and stubble.
• Eye and mouth regions under strong lighting — reflections, shadows, and catchlights in the swapped face should match the scene’s light sources exactly; mismatches are hard to synthesize correctly.

For text-to-video content, faces generally look coherent but the same fine-detail failures that affect still images appear here: overly smooth skin, eyes that are slightly too symmetric, teeth rendered as a uniform block rather than individual structures.

This is the same detection logic applied in face-swap analysis for still images, extended to the temporal dimension.

4. Read on-screen text and fine detail

Generative models have consistently struggled with rendering legible text inside a synthesized scene. Street signs, shop fronts, labels, and subtitles in AI-generated video tend to contain character-like glyphs that are not actual letters, or that shift between frames. This is a high-reliability signal when present.

The same principle applies to other fine detail: number plates, book spines, instrument panels, and any element that requires a specific configuration of small features. If the video contains such elements, zoom in or step frame-by-frame and read them.

In face-swap deepfakes this signal is less relevant (the background is real footage), but text in documents visible in the scene — ID cards, name badges, certificates — may be corrupted if the model partially hallucinated the scene.

5. Look for boundary blur and the subject-background seam

Both video families produce boundary artifacts, though for different reasons.

Text-to-video generators allocate model capacity to the main subject and render backgrounds with less fidelity. The boundary between the subject and the background is a zone of frequent artifact: soft halos, uneven edge sharpness, or abrupt transitions in blur level that don’t correspond to natural depth-of-field.

Face-swap deepfakes show compositing seams at the face boundary that vary frame by frame — the blending algorithm must re-run each frame, and the result is not perfectly stable.

Zoom into the subject-background boundary and step through frames slowly. Consistency in the boundary region across frames is a property of genuine footage; instability there is a flag.

How FakeRadar analyzes video

Manual frame-stepping surfaces obvious cases, but it does not scale to longer clips and it cannot read the signals that live below the visual surface. FakeRadar’s video analysis pipeline addresses this:

1. Frame extraction — the clip is sampled into keyframes distributed across its duration.
2. Per-frame multi-engine scoring — each extracted frame is run through the same AI detection stack as still images: Hive AI model score plus Sightengine for Pro users.
3. Per-frame FFT analysis — frequency-domain analysis on each frame surfaces generation fingerprints (the periodic grid artifacts common in diffusion-model outputs) that are invisible to normal viewing.
4. Per-frame ELA — error-level analysis detects compression-level inconsistencies between regions, which face-swap compositing creates even after re-encoding.
5. Temporal aggregation — results across frames are aggregated so you can see whether signals are consistent (genuine content) or erratic (generation artifacts that vary by frame).

The output reports signals detected across frames, not a single binary verdict. A high-confidence result on a majority of frames means something different from one frame flagging while others do not. Reading the pattern is part of reading the result — which is why the signal-based approach matters as much here as it does for images.

For a deeper look at what per-frame frequency analysis reveals, the deepfake job interview detection article covers a face-swap video case in detail.

FAQ

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

Sometimes, but not reliably. The remaining tells — temporal flicker, physics failures, boundary blur — are subtle and disappear after compression. Frame-by-frame stepping catches what real-time viewing misses.

What is the difference between a deepfake video and an AI-generated video?

Text-to-video generators synthesize entirely new scenes from a prompt. Deepfake videos replace one face in existing footage. Both leave forensic traces but the artifact profiles differ — physics and texture for text-to-video, boundary seams and skin inconsistency for face-swap.

Does FakeRadar analyze video audio for deepfake detection?

No. FakeRadar’s video analysis is visual only. You can observe lip-sync mismatch as a visual flag, but the audio stream is not analyzed.

How many frames does FakeRadar analyze in a video?

FakeRadar samples keyframes across the clip and runs AI detection, FFT, and ELA on each. This catches inconsistencies that a single-frame or thumbnail check would miss entirely.

Summary

• AI-generated video requires frame-by-frame inspection, not real-time watching.
• Temporal flicker on stationary elements is the strongest single class of text-to-video artifact.
• Face-swap deepfakes leave boundary seams and skin-texture inconsistencies at the face compositing region.
• On-screen text corruption, physics failures, and subject-background boundary instability are all reliable flags.
• Visual lip-sync mismatch is observable but audio analysis is outside FakeRadar’s scope.
• Multi-engine frame-level analysis surfaces generation fingerprints that are invisible to normal viewing.
• Analyze any video clip for free when a frame-by-frame review leaves you unsure.