Google has expanded its SynthID tool for detecting AI-generated or AI-edited content and launched a dedicated website that allows anyone to check images, video, and audio for an invisible watermark. Instead of relying on Gemini for an answer about whether the watermark is present, users can now upload the file to SynthID.com and receive a direct result.
What has changed?
SynthID embeds a hidden signal within the pixels of images and video or within the audio waveform. Google says Gemini has integrated the technology into more than 180 billion images and videos, in addition to the equivalent of 240,000 years of audio content. Previously, checks outside the trusted tester program went through Gemini, an approach that was not always practical, particularly when users needed a clear yes-or-no answer.
The biggest limitation of Google’s previous tool was that it checked only Google’s version of SynthID. The new website supports the different watermarks used by SynthID partners, including OpenAI, Nvidia, and Kakao, with Apple expected to join, according to Google’s announcement. This makes the website a single checking point instead of requiring separate tools for each provider.
Usage limits and results
The website does not display the details that were available in the internal tool, such as identifying the areas containing SynthID pixels within an image. Instead, it tells users only whether the watermark was used to generate or modify the media.
Using the service requires signing in, and it accepts Google, OpenAI, and Apple accounts. Google also imposes a daily quota of approximately 10 image, video, and audio checks per user. According to what Ars Technica reported from Google engineers, this restriction is intended to limit the use of the tool to develop ways to remove or bypass SynthID. Actual access may decrease when very similar files are checked, as this may indicate an attempt to tune a workflow to bypass the detector.
Why SynthID alone is not enough
Expanding access to SynthID makes it easier to verify the fingerprints of several AI tools, but it does not turn the tool into a comprehensive detector. Some companies, such as Meta, use different labeling technologies, while other models do not add any watermark. Open-weight models can also generate content locally without a simple watermark or reliable metadata.
Watermark removal remains possible, meaning that a lack of a SynthID result does not prove that content is original. The editorial reading here is that the launch’s value lies in consolidating and expanding checks to multiple partners, not in definitively solving the problem of verifying content authenticity. The opposite approach, which proves the authenticity of images, video, and audio at the time of creation using cryptographic technologies such as C2PA, offers a different path; the article mentions Pixel phones among the devices that implement it.