I examined Google’s new SynthID detector on AI slop. There’s a giant catch

I plopped the ultra-cute video into the device, waited a couple of seconds, and received the outcome: no AI watermarks detected. So, the video have to be actual, proper?

Properly, not precisely.

Developed by Google as a technique of tagging AI-generated content material, SynthID permits AI fashions to embed invisible watermarks within the photos, movies, audio information, and different content material that they create. There’s even a model of SynthID for watermarking textual content, however this specific SynthID detection device (which was beforehand obtainable to researchers, journalists, and different media professionals) is proscribed to picture, video, and audio codecs.

In fact, not each AI mannequin helps SynthID watermarking, and for now, Google’s SynthID device can solely detect AI-generated media created by fashions from Nvidia, Google, OpenAI, and South Korea-based Kakao.

Put all that collectively, and it means Google’s SynthID detection device can inform you one in every of two issues: that it detected a SynthID watermark in a given piece of content material, or that it didn’t detect one. That’s all.

To place the device to the check, I requested Gemini and ChatGPT to create some AI slop for me — specifically, photos and movies of “a robotic visiting Disneyland.”