The full video is uploaded to Gemini 2.5 Flash, which watches it like a human reviewer — checking for unnatural motion, temporal flicker, inconsistent lighting, and AI-specific artifacts.
6 keyframes are extracted and sent to Gemini individually. Each frame is checked for skin smoothness, hair coherence, hand anatomy, and face boundary blending.
ffmpeg runs five low-level analyses — PSNR inter-frame comparison, YDIF luma variance, freeze detection, motion vectors, and codec/bitrate anomaly checks. No external API needed.
A dedicated pass targeting AI avatar tools. Looks for the hallmark pattern: fully static background (YDIF ≈ 0) with a moving, speaking face overlaid. Jewelry and hair that don't move are also flagged.
If Avatar Detector or Pixel Forensics fire strongly, they override a REAL verdict from Gemini — this catches HeyGen-style videos that fool visual analysis alone.
AI GENERATED
≥ 85
LIKELY AI
55–84
UNCERTAIN
35–54
LIKELY REAL
15–34
REAL
< 15
Works with YouTube, TikTok, Instagram links and direct video file URLs (.mp4, .mov, .webm).
Demo. This page shows the tool's interface and how it scores what it finds. Live analysis is not running for you yet.
Live access is by invitation. Verify an approved email to run this page against your own target, or ask for access.
AI detection is probabilistic — results are not guaranteed. Videos are sent to Google's servers via Gemini AI and deleted immediately after analysis.