How to Tell if a Video Is AI-Generated: A 2026 Field Guide

Spotting AI-generated video in 2026 is meaningfully harder than it was two years ago. Sora, Kling, Runway Gen-3, and Pika 2.0 have each addressed the most obvious visual artifacts that made early AI video easy to dismiss. The camera grain is realistic. The motion is smoother. Human faces no longer shatter into abstract geometry during fast movements. But the tells haven't disappeared — they've migrated. This guide covers where to look now, how to use tools when your eyes aren't enough, and how to build a verification habit that keeps pace with the technology.

Why 2026 Is Different

The video generation landscape changed significantly in 2024 and 2025. First-generation tools like early Runway and Stable Video Diffusion produced video that flickered, distorted faces, and generated physically impossible motion. The new generation — Sora, Kling 2.0, Runway Gen-3, Pika 2.0 — produces output that passes casual inspection. This matters because it changes the burden of proof: you can no longer rely on a single jarring artifact to reach a verdict. Verification now requires systematic examination across multiple dimensions, and often a forensic tool for final confirmation.

Temporal Coherence: The Most Reliable Tell

The single most consistent artifact in AI video is temporal incoherence — frame-to-frame inconsistency in details that should remain stable. Watch a suspicious video at 0.25x speed and look for: skin texture that smooths then re-sharpens between frames; background elements that shift position slightly from one frame to the next; hair strands that change their path across frames; and fabric folds that jump rather than flow. These micro-inconsistencies occur because AI video models generate frames with some degree of independence, and the continuity constraints, while improved, are not perfect. Human-made video is physically consistent because it captures the same real scene from frame to frame.

Physics Artifacts: Hair, Water, Cloth

Physical simulation is where current AI video generators still struggle most visibly. Hair is particularly revealing — real hair moves as a unified system under gravity, wind, and momentum. AI-generated hair often produces individual strands that move independently or disappear between frames. Water in AI video frequently lacks the surface tension and light behavior of real water — it tends to look slightly too smooth or to animate with a looping, repetitive quality. Cloth should drape and move according to the physics of the fabric and the body beneath it; AI cloth sometimes flows in directions that don't correspond to the character's motion or to any apparent wind source. When a video contains any of these elements, slow it down and watch them carefully.

Background Inconsistencies

AI video generation has a tendency to treat the background as less important than the foreground subject. This produces backgrounds that are subtly inconsistent from shot to shot — not wrong enough to be immediately obvious, but off enough to flag under scrutiny. Look for: architectural elements that shift between shots (a window that's in a slightly different position, a wall corner that changes angle); crowd scenes where different people appear to be wearing slightly different clothes between cuts; and text on signs or screens that changes content or formatting across frames. Real video, even with cuts and multiple takes, captures a consistent physical environment.

Camera Grain and Noise Patterns

A reliable but less intuitive tell is the noise pattern of the video. Real cameras produce a characteristic pattern of grain or sensor noise that is consistent and random in a way that reflects the sensor physics. AI-generated video either lacks grain entirely (producing unnaturally clean imagery) or adds grain as a post-processing step that doesn't match how real camera noise behaves — it may repeat, appear in unnatural patterns, or fail to correlate with the apparent lighting conditions. Zoom into a still frame from a suspicious video and examine smooth surfaces (sky, walls, skin) for noise texture. AI video noise is often subtly wrong.

rPPG Analysis: The Heartbeat Test

Remote photoplethysmography (rPPG) is a technique that detects blood flow in facial skin by analyzing subtle color changes caused by the heartbeat. Real human faces in video exhibit these changes even when invisible to the naked eye. AI-generated faces typically do not, because the generation model doesn't simulate physiological processes. Forensic tools that incorporate rPPG analysis can flag synthetic faces even when they appear visually convincing. This is one of the more powerful detection methods for close-up talking-head style AI video, which is common in deepfakes and AI-generated news anchors.

Tool-Specific Tells in 2026

Each major AI video generator leaves its own fingerprints. Sora tends to produce very high production-quality video with cinematic lighting but occasionally generates environments that are almost too perfect — backgrounds that look like CG renders rather than captured reality, and motion that lacks the imperfections of handheld or even tripod-mounted cameras. Kling produces excellent facial detail but sometimes struggles with long limbs and peripheral body parts — watch the hands and feet when they appear near the edges of frame. Runway Gen-3 handles dynamic scenes well but can produce transitions between motion states (stop to start, change of direction) that feel slightly mechanical. Pika is strongest on short clips but shows temporal artifacts on longer generations. None of these are absolute rules, but they're useful starting hypotheses.

When to Use Forensic Tools

Your eyes are useful for initial triage but insufficient for high-stakes decisions. Forensic video analysis tools examine: frequency-domain patterns in individual frames (the same approach used for image detection); temporal consistency metrics across frame sequences; audio-visual synchronization analysis; and noise distribution patterns that differ between real camera capture and AI generation. SUS IT's video analysis pipeline applies these methods to flag suspicious content even when it passes visual inspection. For any video where the stakes matter — journalism, legal evidence, financial decision-making — run it through a forensic tool rather than relying on visual judgment alone.

A Practical Verification Checklist

When you encounter a suspicious video: (1) Slow it down to 0.25x or 0.1x and examine frame-by-frame in the most dynamic sections. (2) Check the physics — hair, water, cloth, and peripheral body parts. (3) Examine the background for consistency across cuts. (4) Look at noise patterns in smooth surfaces zoomed in. (5) Watch the edges of faces and hair against backgrounds for blending artifacts. (6) Check lip sync carefully on any speech — advance or delay is still common. (7) Run it through forensic tools before making any consequential judgment. A single artifact isn't proof; a pattern of artifacts is. When you can't find artifacts but the video's content is extraordinary or the source is unverified, treat it as unverified regardless of what your eyes tell you.

The Limits of Verification

No verification method is perfect. Heavily compressed or re-encoded video loses the high-frequency information that forensic tools depend on. Future AI video generators will fix many of the tells described here. The right posture isn't "I can always detect AI video" — it's "I apply systematic scrutiny before treating extraordinary video as genuine." In a high-quality deepfake, uncertainty is the honest conclusion. The question is whether you communicate that uncertainty or pretend to certainty you don't have.