Google Launches SynthID Website for Global Detection
Pushmeet Kohli, Vice President of Science and Strategy at Google DeepMind, announced that the SynthID Detector has expanded from a trial version for media professionals to a global audience. The website is now available worldwide, with an English interface. Users can upload images, videos, or audio to check for AI watermarks from Google or its partners.
SynthID was launched in 2023, embedding invisible signals into image and video pixels, as well as audio waveforms. Google claims to have watermarked over 18 billion images and videos, along with audio equivalent to 240,000 years. Media generated by Gemini, Flow, ProducerAI, Vids, as well as Nano Banana, Veo, and Lyria will carry this mark. Last year, the detection entry was first opened to journalists, media, and researchers, with the public previously only able to inquire within Gemini, where the model sometimes could not provide a clear yes or no.
The new site removes Gemini from the detection process. Ars Technica reports the address is SynthID.com. It no longer only recognizes Google's own watermark version but also acknowledges the watermarks of partners, including OpenAI, Nvidia, Kakao, with Apple set to join soon. TechCrunch lists supported image formats as JPG, PNG, WEBP, GIF, TIFF, with video formats including MP4, MOV, WEBM, and audio formats as WAV, MP3, FLAC, AAC, etc. OpenAI also has a detection site but originally only recognized its own signature.
Detection is also embedded in search, the Gemini app, and Chrome. Google states these built-in verifications now handle over 1 million requests daily. The website adds the same context, not a new content removal right. No watermark does not mean the file is not generated. Models, screenshots, recompressions, and crops that have not adopted this watermark are not covered by this announcement.
Buyers include newsrooms, platforms, and ordinary uploaders who need to confirm the source of materials before publication, while sellers are model companies that still issue their own watermarks and have non-interoperable detection points. This is a verification flow driven by the opening of detection entry, not a previously announced authentication fee. Requests will shift from "asking Gemini once" to a dedicated website, along with over 1 million daily verifications in search and browsers. Beneficiaries include Google, OpenAI, Nvidia, and Kakao that have integrated SynthID, while those without this marking face pressure.
Microsoft and Meta use a different set of watermarking and verification standards. Text watermarks were previously open-sourced, and this announcement focuses on images, videos, and audio for the general public.
Source: Public Information
ABAB AI Insight
In 2023, Google developed SynthID to embed signals during generation rather than relying on visual guesses afterward. In May 2025, the detection entry was first provided to journalists and researchers, with over 10 billion pieces of content already marked at that time. By October 2026, the number increased to over 18 billion images and videos, plus audio equivalent to 240,000 years. The strategy is to first secure the exits of its own Gemini, Veo, and Lyria, then allow OpenAI, Nvidia, and Kakao to send their respective watermarks into the same detection point.
The traffic has not yet turned into authentication fees. The verifications in search, Gemini, and Chrome often exceed 1 million requests daily; the website simply breaks down the previous step that required entering a dialogue to ask. OpenAI still retains its site that only recognizes its own signature, while Apple is noted as soon to join. The financial and distribution aspects remain at the default exit: only those whose models embed marks during generation will receive affirmative answers on this site.
This is at the same stage as camera manufacturers handing original image verification to C2PA, with newsrooms using original documents for verification. The difference is that SynthID embeds marks into pixels and waveforms, without relying on users to save a separate proof. The industry is concentrated at the entry point, not at the completion of detection: Microsoft and Meta have their own standards, and models, screenshots, and recompressions without watermarks can still bypass. The 2025 trial entry only recognizes Google, while the 2026 public site will begin to recognize partners.
Essentially, this represents a shift in pricing power. The pricing for content authenticity shifts from "does it look generated" to "was a verifiable mark embedded during generation." The mechanism is that once the detection point is centralized, it becomes harder to prove without a mark. Google leaves a verification position with millions of requests daily, while model companies must decide whether they are willing to embed the same signal at the exit.
ABAB News · Cognitive Law
If no mark is embedded during generation, post-detection can only deny.
Once the detection point is centralized, tools that are not integrated lose proof first.
A watermark is not the authenticity itself; it is a receipt left by the exit.