Google has officially launched a standalone public-facing verification website, synthid.com, where anyone can upload image, video, or audio files to detect whether they carry a SynthID digital watermark.
SynthID was developed by Google DeepMind. It first appeared in August 2023 alongside Imagen, and then gradually expanded to audio, video, and text generation.
SynthID's principle is not to attach a visible badge, nor to write into metadata that can easily be stripped, but to embed the identification signal directly into the content's underlying layer. For example, it makes imperceptible fine adjustments to pixels and audio spectra, and for text generation it fine-tunes token sampling probabilities.
At present, the entire product lineup, including Gemini, Imagen, and Veo, has deeply integrated it. Chrome and Search also have built-in verification, and you can also verify directly with Gemini and Antigravity.
SynthID's ecosystem has been expanding continuously. In addition to OpenAI, NVIDIA, and Kakao having already integrated it, Apple is also about to follow suit. Official data shows that the system has cumulatively marked 180 billion images and videos and more than 240,000 years of audio, with daily verifications reaching one million.
One additional note here: the above content is basically the official line. The actual situation is that the watermark Claude launched a while ago uses SynthID-Text, while the textGrain watermark OpenAI launched a few days ago is self-developed. The reason it has been included in official deep cooperation is mainly that its images and other content use SynthID.
The shortcomings of textGrain were mentioned earlier. SynthID is not invulnerable either. It can withstand everyday cropping, compression, and filters, but when faced with semantic rewriting, multiple translations, adversarial perturbations, or aggressive redrawing, the hidden signal may still be erased.
Uh, isn't that just stating the obvious? If you do all that, even an ironclad watermark would be worn away. The main thing is that most people have no energy to deliberately erase watermarks after creating with AI, so in most scenarios verification is still possible.
Personally, I think that as AI content becomes increasingly realistic, this kind of watermark verification is very necessary.
Now when you open various social media apps, unmarked AI content can really fool a bunch of people. Especially the old products of tabloid literature: back then they relied entirely on text descriptions, but now they can be generated directly with AI, repackaged and brought out to fool people again.