Google Opens SynthID Detector to Verify AI Content Globally

Google DeepMind is opening its SynthID Detector to everyone. Starting today, the verification tool is available globally in English, letting anyone check whether an image, video, or audio file was created or edited with AI. Until now the detector was restricted to a hand-picked group of journalists, media professionals, and researchers who tested it after Google I/O last year.

What happened

The new detector, reachable through a dedicated website, no longer stops at Google's own watermark. Google says it now reads the distinct SynthID signals of its partners too, including OpenAI, Nvidia, Kakao, and, soon, Apple. That fixes the awkward part of the old tool: an AI image made with ChatGPT already carried SynthID, but Google's portal would not flag it unless it had been produced by Gemini or Veo.

Writing on the Google blog, DeepMind VP of science Pushmeet Kohli said the company has watermarked more than 180 billion images and videos, along with 240,000 years of audio. Google says people already make roughly a million verification requests a day, and SynthID detection is also baked into the Gemini app and Chrome.

Access is open but not unlimited. You must log in with a Google, OpenAI, or Apple account, and each user gets a daily quota of about ten image, video, and audio checks. The site accepts most common formats, from JPEG, PNG, and HEIC to MP4, WEBM, WAV, and MP3. Unlike the internal tool, it returns a simple yes-or-no verdict rather than highlighting the regions that carry the mark.

Why it matters

Content provenance is becoming a normal layer of the internet, and Google is trying to make its standard the default place to check it. The timing is not incidental: the EU AI Act's transparency obligations push companies to label synthetic media, so a free, public detector gives creators and auditors a concrete way to comply.

The limits matter just as much. Google's partners are only part of the AI ecosystem. Meta relies on its own labeling scheme, plenty of providers use none at all, and open-weight models run on personal hardware produce images with no watermark to find. Google's own engineers told Ars Technica that the daily quota exists to stop people tuning bypass tools, and the company concedes that some AI content will always pass the check unflagged. As TechCrunch notes, the tool is a useful single destination, not a guarantee.

What this means for builders

• Provenance checking is now a pipeline step, not a research project. If your agents generate or publish media, a verification call belongs next to your logging and moderation hooks.
• Watermarks are not proof of authenticity. Plan for content that carries no signal at all, especially anything coming from open-weight models or Meta tools.
• Compliance is a moving target. The EU's labeling rules are pushing watermarks and detectors into default workflows, so treat provenance metadata, such as C2PA, as part of your data model.
• Trust features sell. As models converge on quality, verifiable output is becoming a differentiator buyers actually ask about.