Google now lets anyone check if an image, video, or audio clip was made with its AI tools, and the checker mostly does its job.
The company launched a standalone web tool for its SynthID watermark, letting users upload files and get a read on whether Gemini, ChatGPT, or another SynthID-partnered tool produced them. A tester ran dozens of images through it: straight AI outputs were flagged instantly, screenshots and oil-painting filters didn't fool it, and even a photo stitched together from two different AI tools got caught. The checker broke only twice, once after heavy pixelation and once on a composite that was 75% real photo and 25% AI. Oddly, the detector still caught a separate image where ChatGPT had added just one extra person to an otherwise real group photo, despite that image containing even less AI content than the composite that fooled it.
That inconsistency is the real finding here. If detection depended purely on how much of an image is AI-generated, the single-person insert should have been the easier file to sneak past, not the harder one. It suggests SynthID's reliability hinges on how the AI content was generated and blended into a scene, not just its share of the final image, which makes the tool harder to game than a simple percentage rule, but also harder to predict.
Watermarking is still a patch, not a cure. OpenAI runs its own separate detector, and Google caps this one at roughly ten files a day, so neither company is pretending this scales to policing the internet's AI output.