Google has opened the SynthID Detector to anyone, in English and at global scale, and it now identifies watermarks beyond its own models. Announced on October 7, 2026 by DeepMind, the SynthID Detector checks whether an image, video or audio file was produced or edited with AI from Google or from partners, including OpenAI, NVIDIA and Kakao. Apple is named for a later stage.
The post is by Pushmeet Kohli, vice president of science and strategic initiatives at Google DeepMind. The detector lives at synthid.com. Until now, outside a program limited to media professionals, the usual path was to ask Gemini whether a file carried a SynthID mark. The new site takes the chat model out of the loop and returns a direct answer about the mark.
What the mark covers, and what it does not
SynthID has existed since 2023 as an imperceptible signal embedded in image and video pixels and in the audio waveform. Google says it has watermarked more than 180 billion images and videos and the equivalent of 240,000 years of audio. The earlier detector only read Google's own mark. ChatGPT images, for example, could carry SynthID and still fail the check. That limit no longer applies to partners already integrated.
The tool joins verification already built into Search, the Gemini app and Chrome, which Google says regularly handle more than 1 million requests a day. It is not a universal authenticity test. Ars Technica notes that the site should make the limit clear: models from other companies, such as Meta, use different technology or no watermark, and open models run locally can generate media without the signal. A negative result does not prove a file is real.
Login, quota and the bypass risk
Use requires a login with a Google, OpenAI or Apple account. According to Ars Technica, citing Google, each user gets a daily quota of about 10 image, video or audio checks. Google engineers told the outlet that the limit exists to make it harder to use the queries to build tools that strip SynthID. The public detector is also simpler than the internal version: it reports whether the mark was found, without highlighting the regions where the signal appears.
For anyone tracking deepfakes, generated art and the origin of videos circulating online, the change is practical: one address now reads marks from several AI companies instead of one detector per maker. What remains open is Apple's participation, reach beyond English, and whether the mark survives aggressive edits. Google treats SynthID as context, not as definitive proof.
Sources
- Google DeepMind: SynthID Detector expansion
- SynthID Detector
- Ars Technica: global detector and its limits
Transparency: This content was created, edited or reviewed with the aid of artificial intelligence. The information was cross-checked with public posts on X and sources available on the internet. Consult the original sources for the full context.
Por GeekikiBot