Tired of fake content online? The tool that will help you filter it

Google DeepMind is opening its SynthID Detector tool for free public use, allowing anyone to check whether images, videos, and audio files were created by artificial intelligence. The tool scans content from models by OpenAI, Nvidia, Kakao, and soon Apple, offering a coverage advantage over competitors like Anthropic. However, the technology has limitations: aggressive cropping, extreme compression, or translation can erase the watermarks.

Google DeepMind is opening its SynthID Detector tool for free public use, allowing anyone to upload image, video, and audio files and check whether they were created by artificial intelligence. The tool, previously available only to journalists, scans content not only from Google's Gemini but also from models by OpenAI, Nvidia, Kakao, and soon Apple. The main advantage is its broad coverage compared to OpenAI's Verify tool, which is limited to detecting content from ChatGPT alone. On the other hand, Anthropic does not offer a public tool but a closed API interface for researchers. Lab tests show 97%-99% accuracy in identifying original files, with a false positive rate below one percent. The watermark survives compression and color changes with 85%-95% accuracy. However, aggressive cropping of more than 20% of the image edges or extreme compression can erase the stamp. In text, light human editing and replacing a quarter of the words reduce detection chances to just 17%, and translation completely erases the watermarks. Experts emphasize that no tool is a 'general AI detector' but rather a source verification mechanism that detects specific pre-embedded stamps. A negative result does not prove the content is authentic.