Google Launches Global SynthID Detector for AI Watermarks
Google launched SynthID.com, a public site detecting AI watermarks in images, video, and audio from Google, OpenAI, Nvidia, Kakao, and soon Apple. Previously limited to trusted testers, the tool offers yes/no results after signing in with a Google, OpenAI, or Apple account. Daily quota is ten checks to prevent abuse. The detector misses Meta's watermarks, unwatermarked models, and local open-weight models. Google's Gemini models have watermarked 180 billion images/videos and 240,000 ye
GGLOBAITOOLS DESKSHARE
Google launched SynthID.com, a public site detecting AI watermarks in images, video, and audio from Google, OpenAI, Nvidia, Kakao,…
Share this post
Short answer: Google launched SynthID.com, a public site detecting AI watermarks in images, video, and audio from Google, OpenAI, Nvidia, Kakao, and soon Apple. Previously limited to trusted testers, the tool offers yes/no results after signing in with a Google, OpenAI, or Apple account. Daily quota is ten checks to prevent abuse. The detector misses Meta's watermarks, unwatermarked models, and local open-weight models. Google's Gemini models have watermarked 180 billion images/videos and 240,000 years of audio.
Google SynthID detector public launch
Google has opened its SynthID detection system to the public through a new dedicated website, letting anyone check images, videos, and audio for invisible watermarks from multiple AI providers. The rollout began on October 7, 2026, and marks the first time the tool has been available outside Google's trusted tester program. Until now, the only way to scan for SynthID labels was to ask a Gemini model directly, a process that often returned confusing results when users just wanted a straightforward answer.
SynthID embeds an imperceptible signal into the pixels of AI-generated images and video frames, as well as into the waveform of synthetic audio. Google says its Gemini models alone have already stamped more than 180 billion images and videos with these markers, along with 240,000 years' worth of audio. The technology has also gained adoption from other major players in the field, but verifying those watermarks used to require jumping between different systems or relying on Google's internal tools, which highlighted specific regions where the signal appeared.
SynthID supports partner watermark formats
The new SynthID.com site changes that by supporting the distinct watermark formats used by all of Google's announced partners. That list includes OpenAI, Nvidia, Kakao, and, in the near future, Apple. A single upload now checks for signatures from any of those systems and returns a simple yes-or-no message rather than the detailed heatmaps the internal version provides. The trade-off is clarity over granularity, but for most practical purposes the binary result is what matters.
Access requires signing in with a Google, OpenAI, or Apple account, and each user receives a daily quota of roughly ten checks across images, video, and audio combined. Google engineers have explained that the limit exists to prevent bad actors from using the detector as an oracle while they develop tools to strip or evade SynthID. The system may also cut off access earlier if it detects patterns such as repeated submissions of nearly identical files, which could indicate someone is probing for weaknesses.
SynthID detector limitations and blind spots
Even with the expanded coverage, the detector has blind spots. Meta uses its own watermarking approach, other providers apply no watermarks at all, and open-weight models running on local hardware can produce convincing synthetic media without any embedded signal. The website will carry a disclaimer making this clear. It is also possible that determined researchers will eventually find ways to remove SynthID without degrading the content, which would let watermarked material pass as authentic.
Those gaps highlight why some experts favor the opposite strategy: cryptographically signing real media at the point of capture. The C2PA standard does exactly that, and Google's Pixel phones are among the few devices already implementing it. Until provenance-based verification becomes widespread, however, SynthID remains the most practical option for spotting AI-generated fakes from the largest commercial models.
SynthID API for developers and moderation
For developers and product teams, the launch means there is now a free, centralized endpoint to integrate into content moderation pipelines or user-facing features. It is worth building a fallback that flags the detector's limitations rather than treating a clean result as proof of authenticity. Users who encounter suspicious media should run it through the site but remember that a negative result does not guarantee human origin. The tool is a useful layer in a broader verification strategy, not a silver bullet.
Frequently asked questions
When did Google launch the public SynthID detector website?
Google launched the public SynthID detector at SynthID.com on October 7, 2026. This marked the first time the tool was available outside Google's trusted tester program, where previously the only way to scan for SynthID watermarks was by asking a Gemini model directly.
Which companies' AI watermarks can the new SynthID detector identify?
The detector checks for watermarks from Google and its announced partners: OpenAI, Nvidia, Kakao, and, in the near future, Apple. A single upload scans for signatures from all supported systems and returns a simple yes-or-no result instead of the detailed heatmaps provided by Google's internal tools.
What are the usage limits for the public SynthID detector?
Users must sign in with a Google, OpenAI, or Apple account and receive a daily quota of roughly ten checks across images, video, and audio combined. Google enforces this limit to prevent bad actors from using the detector as an oracle to develop watermark-removal tools, and may cut off access earlier if it detects probing patterns like repeated submissions of nearly identical files.
What are the main limitations of the SynthID detector?
The detector cannot identify watermarks from Meta, which uses its own system, or from providers that apply no watermarks at all. Open-weight models running locally can also generate synthetic media without embedded signals. The site displays a disclaimer about these gaps, and researchers may eventually find ways to remove SynthID without degrading content, letting watermarked material pass as authentic.
How should developers and users treat results from the SynthID detector?
Developers should integrate the free endpoint into moderation pipelines but build fallbacks that flag the detector's limitations rather than treating a clean result as proof of authenticity. Users should run suspicious media through the site but remember a negative result does not guarantee human origin. The tool is a useful layer in a broader verification strategy, not a silver bullet.
Google DeepMind, Meta, and Isomorphic Labs have pledged a combined $300 million to Biohub, the nonprofit founded by Mark Zuckerberg and Priscilla Chan, to build a virtual cell, a computational model designed to simulate biology and predict experimental outcomes digitally. The investment anchors a broader $1.8 billion Virtual Biology initiative matched by over $1 billion in U.S. Department of Energy and NIH commitments.
Liquid AI released two open-weight decision models-d1-3B and d1-omni-600M-on October 7, 2026, targeting edge hardware with single-forward-pass latency. Unlike generative LLMs, they output structured answers (yes/no, categories, scores) in milliseconds: d1-3B hits 16 ms on a Jetson AGX Thor and leads sub-10B models on the Decision Index at 48.57. Both are on Hugging Face with a demo arcade; d1-omni-600M is an early research release handling text-plus-image or text-plus-audio inputs.
French AI lab Mistral released Mistral Large 4, an open-weight model claiming parity with top closed proprietary systems. Launched in preview October 7, 2026, it was trained from scratch and excels in coding, cyberdefense, manufacturing, and finance. Mistral charges only for compute, not model access. The release follows U.S. restrictions on proprietary models and guardrail failures, boosting demand for open alternatives. Mistral recently raised $3.3 billion at a $24 billion valuation.
NO COMMENTS YET
Comments are open. Have a thought or a question? Share it below.