Google launches SynthID Detector to spot AI-generated content, but with significant limitations

Google has launched SynthID Detector, a web service open to any user that can identify AI-generated images, video and audio. According to SiliconANGLE, the tool recognizes content generated with models from Google, OpenAI, Nvidia and Kakao. Using it requires signing in with a Google, OpenAI or Apple account. In a few weeks it will be expanded to also identify images created with Apple's AI tools, such as Image Playground.

SynthID is an imperceptible watermark that Google developed in 2023 and that is added to content generated with the tools of the companies mentioned. Until now, detecting it required special tools. Google had offered various versions of its detector to journalists and other media professionals, but with very limited availability. The company already offered other ways to check, such as the Circle to Search feature on phones or the option to upload content to Gemini and ask whether it is artificial. The new detector makes them more accessible.

The limitations are significant. Because the tool looks only for SynthID watermarks, it performs poorly with AI content that lacks them, and there is a lot of it. According to the article, Anthropic's Claude does not use SynthID, nor does xAI's Grok chatbot, nor do the many Chinese open-weight models. As a result, the detector will miss a large amount of synthetic material.

Another limitation is that it does not distinguish between content created entirely with AI and content merely edited with it. In the latter case, it cannot indicate which part is fake and which is real.

The article concludes that it is a step in the right direction but not enough: there is still no foolproof method for verifying AI-generated content, partly because of the multitude of watermarking standards. Meta, for example, has experimented with its own tools and has a system for identifying images created with Muse Image, which does not work with images from third-party models. According to the author, without a universal standard accepted by the major developers, including Chinese ones, there will probably never be a reliable way to detect this material. Google, given its influence and the partners it already has, would be well placed to drive it.

Google acknowledges the problem on the detector's website: it says it is continually working to expand its partnerships to detect more types of AI-generated media. It adds that building industry standards takes time and that each type of content (image, audio, video) has different technical requirements. Its goal, it says, is a layered ecosystem in which watermarks, metadata standards and verification tools work together across the industry.