Google releases a free tool to check for an increasingly common AI watermark

Google has made its free tool for identifying SynthID watermarks available to the general public, potentially making it easier to tell whether what you’re looking at was created by or edited with generative AI. While the tool is only looking for a watermark, rather than analyzing the image to see if it’s AI-generated, an increasing number of companies are starting to add SynthIDs to the images, video and audio generated by their AI tools.

The tool, called SynthID Detector, is relatively simple. You can upload an image, video file or audio clip to it, and it will scan for a SynthID. Google says that AI tools from itself, Nvidia, OpenAI and Kakao use the standard, so the tool should be able to tell whether AI was used to create the thing you’ve uploaded. Apple has also promised to start using SynthIDs to watermark its AI-powered photo manipulation tools later this year.

This image was originally a black and white film photo, but was colorized using Google Gemini, which added a SynthID watermark to its output. The tool was then able to detect it. Mitchell Clark

Rather than try to add easily-removed metadata, SynthIDs are invisible watermarks added to AI-generated media, which should theoretically survive modification like cropping, adjusting colors or being compressed by a platform that isn’t even attempting to preserve the watermark. Also importantly, the watermark shouldn’t be detectable in the final output.

There is one somewhat annoying gotcha with the tool: it requires you to sign in with a Google, Apple or OpenAI account to actually check anything. This extra friction may be enough to dissuade someone from casually checking an image, and doesn’t really feel like it should be necessary.

It’s also worth noting that it is very limited. While many major players are represented in the list of companies on board with SynthID, there are plenty that aren’t. As the tool’s homepage warns, it doesn’t detect whether the image was made by AI, it can only tell you whether it was made with a specific subset of models. It’s also binary, only telling you if an image has a watermark or not. Unlike more robust (and, in some ways, more limited) tools like Apple’s Reference Image system or Coalition for Content Provenance and Authenticity (C2PA) Content Credentials, there’s no way to see the original image, or tell whether the AI simply tweaked a real photo or created an image wholesale.

This image was edited using with Apple Photo’s AI tools, but because those don’t yet bake in SynthIDs, the tool doesn’t detect it. Even if it could, though, it wouldn’t be able to tell me that the image is a real photograph that’s had part of it erased using AI, rather than an entirely synthetic image.Mitchell Clark

Still, it’s an important piece of the puzzle as SynthID is gaining popularity. After all, a watermark isn’t particularly useful if it’s invisible and you have no way of checking it. You’ve been able to ask Google Gemini if an image has a SynthID watermark in it for a while now, and Google’s Chrome browser and Search engine will flag images with them in some cases, but the detector is a much lighter-weight solution, and provides an option for those who may be checking an image because they’re morally opposed to using AI, and who likely wouldn’t view being able to ask an AI chatbot about it as an acceptable option.