How to Tell if a Video Is AI-Generated: 6 Signs That Never Fail

AI-generated video crossed the "you might not notice" line in 2025. In 2026 the checks that used to work — extra fingers, weird eyes — are mostly obsolete. Here are six that still reliably catch synthetic clips.

1. Watch the physics of small things

Related: GPT-5 Is Here: Everything You Need to Know About OpenAI's Most Powerful Model Yet →

AI models learn dominant motion (people walking, wind in leaves) but stumble on small physical objects.

• Ice cubes floating instead of sinking
• Cloth that folds in impossible ways
• Water splashes with no droplets
• A ring that changes finger between frames

Pause every 0.5 seconds and look at what should be interacting with gravity.

2. Check text in the scene

Related: Will AI Coding Agents Replace Developers? We Asked 100 Engineers →

Signage, book covers, name tags, and license plates are where AI still trips. Even the best 2026 models produce text that:

• Warps between frames
• Uses invented characters
• Mixes letter styles mid-word
• Reads correctly at a distance but breaks up close

If you can read a sign clearly and the letters are stable across a full second, that's a real recording indicator (though not a guarantee).

3. Reflections and shadows

Related: The 27 Best AI Tools in 2026 (Tested for 90 Days) →

Real cameras capture reflections that match the source of light. AI often gets:

• Reflections in eyes that don't match the environment
• Shadows going in inconsistent directions
• Glossy surfaces (car paint, water, glasses) with reflections that look painted rather than mirrored

Look for a car window, a puddle, or a glass door in the shot.

4. Ears, teeth, jewelry

Related: ChatGPT vs Claude 4: Which AI Should You Actually Pay For in 2026? →

Faces are much better in 2026, but the edges of faces still leak. Check:

• Earring shape stability across a smile or head turn
• Teeth count and shape when someone talks (AI teeth often shimmer)
• Necklaces that clip through skin
• Glasses arms that disappear behind the ear inconsistently

5. Camera behavior

Related: Google Gemini 3 Ultra Review: Has Google Finally Caught Up? →

Cinematographers do things with camera that models still fake:

• Depth of field drift — real focus pulls hit a target; AI blur "breathes" in strange ways
• Rolling shutter — real fast pans on phones create characteristic warping; AI usually forgets it
• Motion blur on hands but not on the head during fast movement, or vice versa

6. Metadata and audio

If you can access the file itself:

• Check EXIF / metadata (no camera model = suspicious)
• Look at audio waveform: AI clips often have unnaturally clean sound with no room noise
• Ambient sound that doesn't match the environment (busy street with silent background) is a giveaway

Tools that automate this

Public detectors like Deepware, Hive AI's video detector, and Truepic help — but treat their scores as evidence, not proof. A 90% "AI-generated" score with confirming visual tells is strong; a 60% score alone isn't.

What to do when a video matters

For anything with real consequences — news, legal, personal — combine three checks: visual tells, at least one detector, and a reverse-image search on standout frames. That combination catches 99% of current AI video. The remaining 1% needs a professional forensic team, not a checklist.