Summary
Matt Wolfe set out to build a practical AI-generated video detector — a web tool where users paste a social media link and receive a verdict on whether the footage is AI-made. The project uses ChatGPT’s Codex agent for development, Google’s Gemini video understanding model (including a newly released September version with improved video accuracy), Google SynthID watermark detection, and the SiteEngine API as a third-party detection layer. After more than an hour of automated coding, Wolfe had a working interface accepting YouTube, TikTok, Instagram, and X URLs.
Testing against a personal library of known AI-generated clips revealed deep reliability problems. The system correctly identified 23 of 39 AI clips, but when SiteEngine and Gemini disagreed — a common outcome — the app returned “inconclusive” rather than committing to a verdict. A visually obvious AI clip (a woman in an inflatable suit water-jetting) still produced an inconclusive result after nearly 10 hours of additional iteration and prompt engineering.
Wolfe’s honest conclusion is that current automated detection accuracy falls well short of human visual perception, even for obviously synthetic footage. The video serves as a candid field report on the limits of AI video forensics in late 2026, touching on Google SynthID’s scope (only flags Google-originated AI video), Gemini’s new agentic video understanding capabilities, and the fundamental challenge that competing detection signals from different tools produce unresolvable disagreements.
📺 Source: Matt Wolfe · Published September 09, 2026
🏷️ Format: Hands On Build
