Google rolls out improved SynthID AI content detector, now available globally
Ars Technica · LC · trust 55/100
Finding fakes Google rolls out improved SynthID AI content detector, now available globally The new SynthID website can now identify AI content from Google, OpenAI, and more.
15 Credit: Aurich Lawson Credit: Aurich Lawson Text settings Story text Size Small Standard Large Width * Standard Wide Links Standard Orange * Subscribers only Learn more Minimize to nav Every piece of AI content from Google’s Gemini models has a hidden SynthID label, which makes it very difficult to pass the content off as authentic . SynthID has also gotten buy-in from other major AI players, but actually checking for those invisible watermarks has been a convoluted process. It’s going to be a little easier going forward, though. Google has announced that its SynthID detector now supports watermarks from all its partners, and it’s available for anyone to use on a new dedicated website.
SynthID is an invisible signal encoded in the pixels of AI-generated images and videos, and in the waveform of AI audio. Google says Gemini alone has integrated SynthID with over 180 billion images and videos, plus 240,000 years of audio. Until now, the only way for those outside of Google’s trusted tester program to check for SynthID was to ask Gemini. This is rather clunky, and the model’s output could sometimes be confusing when you just wanted a simple “yes” or “no.”
Now you can cut Gemini out of the loop. Starting today, you can go to the new SynthID.com website to check a piece of content for an AI watermark. A major limitation of Google’s previous detector was that it only checked for Google’s version of SynthID. Even though ChatGPT images, for example, have SynthID, they wouldn’t be flagged in Google’s detector. That’s no longer a problem.
Google says that its expanded version of the detector will support the distinct watermarks from all its SynthID partners, including OpenAI, Nvidia, Kakao, and (soon) Apple. So you’ll be able to go there and check a file for the involvement of all those AI systems instead of accessing their detectors individually. The website won’t provide as much information as the internal tool, which highlighted regions of the image that showed SynthID pixels. Instead, you’ll just get a message telling you whether SynthID was used to produce or edit the media.
You have to log in to use the tool, but it supports Google, OpenAI, and Apple accounts. While the detector is available to anyone, the login requirement limits how much you can use it. Google says each user gets “a daily usage quota of approximately 10 image, video, and audio checks.”
The previous detection tools had a similar limit. Google engineers have told Ars that limiting access is necessary to prevent people from using the checks to develop SynthID removal tools. The fuzziness about the exact number is likely because the system may lock you out sooner if you’re checking multiple images that are very similar. This could signal someone is trying to tune a bypass workflow.
Expanding access to SynthID detection is a good thing—it’s a convenient single destination to check media for the signature of popular AI tools. However, Google’s partners are far from the only sources of AI content.
Some model providers use different watermarking tech (like Meta), while others use no watermarks at all. And then there are open-weight models that people can run on their own hardware to produce increasingly realistic AI content that doesn’t even include simple metadata labeling.
That means there is plenty of AI content that will not be identified as AI in the SynthID tool. The website will have a disclaimer saying as much. It’s also plausible that despite Google’s efforts, someone will find a way to remove SynthID, making existing AI content look more legit. It’s not a perfect system, but that’s the reality of relying on proactive watermarking for AI images, which are proliferating across the Internet at an astonishing rate.
The best approach might be the opposite—using cryptographic technologies like C2PA to label real images, video, and audio as they are created. Google’s Pixel phones are among the few devices doing just that. In the meantime, SynthID is the best bet you’ve got to identify AI fakes.
Read the original at Ars Technica →
Open in TruthVane →