Google Puts SynthID Detector in Public Hands, Betting Provenance Beats Detection
Google's SynthID Detector is now public worldwide in English, checking images, video and audio for AI watermarks from OpenAI, NVIDIA, Kakao and Apple soon.
Google has opened its SynthID Detector to the general public worldwide, ending a restricted pilot that had kept the tool in the hands of journalists and researchers. Since October 7, 2026, anyone can upload an image, video or audio file to synthid.com and check whether it carries an invisible watermark added at the moment of generation. The service runs globally but only in English for now.
Google says its watermarks already sit on more than 180 billion images and videos produced by its own models and by partner tools. The SynthID Detector reads signals from Google's Gemini output and from content made with OpenAI, NVIDIA and Kakao systems, with Apple support slated to follow. Verification also surfaces inside Google Search, the Gemini app and Chrome, and the tool fields more than one million checks a day.
The standalone site is the visible end of a wider rollout. Google is placing verification signals inside Search, the Gemini app and Chrome, so provenance information can appear next to a result or a generated asset without the user visiting synthid.com at all. The public portal works as the fallback for files that arrive by email, messaging or download, where no Google surface stands between the content and the person checking it.
What the SynthID Detector does, and does not, confirm
SynthID is a watermarking system rather than a classifier. It embeds a signal that human eyes and ears cannot perceive into AI-generated images, video and audio at the point of creation, and the detector's task is to find that signal in a file and report it. Google's tally of watermarked audio runs past 240,000 years of material.
Because the mark is designed to be imperceptible, it does not depend on a viewer noticing a label, a caption or an overlay. It travels inside the file itself, which is why the same check works across still images, video clips and audio tracks.
The scope stops at three modalities. Text sits outside the check, even though AI-written text travels faster than generated media through email, chat threads and search results. A tool that verifies a video while saying nothing about the paragraph beside it leaves a visible gap in any newsroom's verification workflow.
That distinction shapes how the tool should be used. A positive result is real evidence that a file passed through a supported generator. A negative result means only that the detector found nothing. Files that were cropped, re-encoded or screenshotted can lose the signal, and output from models that never applied SynthID never carried one to begin with.
The portal therefore answers a narrower question than the phrase "AI detector" implies. It confirms provenance where provenance was recorded, and it stays silent everywhere else. Google's baseline numbers point to demand that predates the public launch: one million daily verification requests show that checking a file has already become routine inside Google's own products, which gives the company an operating history to draw on as outside traffic arrives.
The partner list is the real strategic move
Google built SynthID for its own models. The public detector now reads signals from OpenAI, NVIDIA and Kakao as well, and Apple is named as the next addition. That converts a single-vendor utility into a cross-industry verification layer.
| Provider | SynthID watermark support |
|---|---|
| Google (Gemini) | Live |
| OpenAI | Live |
| NVIDIA | Live |
| Kakao | Live |
| Apple | Announced, not yet live |
Apple's pending inclusion carries the most weight for reach. Its hardware sits at the capture end of much of the world's photo and video, so readable Apple watermarks would extend the detector into a second large consumer ecosystem. The coalition still leaves out most open-weight model developers, whose output carries no mark at all.
Joining the group costs the partners something. Each one has to build watermarking into its generation pipeline and accept that an outside company's detector will read its output. What they get in return is a shared answer to the question of what is real, delivered without any single vendor owning the verdict.
Cross-vendor coverage is what makes the tool usable inside a business process. A newsroom, an insurer or a platform trust team cannot reasonably run one detector per model family. A single portal that reads several watermark formats lowers the cost of a provenance check to a file upload.
Where watermarking stops working
Watermarking has a structural weakness: it travels only with content that preserves it. Editing tools, format conversions and social platforms that reprocess uploads can strip the signal before a detector ever sees it. The detector cannot recover what a pipeline removed.
A second limit is adoption. Google's coverage spans its own products and three named partners. Models outside that group generate content with no SynthID mark, and the SynthID Detector has no way to distinguish that content from a human-made file.
The gap is sharpest for open-weight models. Their developers distribute weights rather than operate a service, so there is no obvious point at which a watermark could be embedded, and their output reaches the detector with no signal attached. For a verification tool, that blind spot is structural.
Google has not published a false-positive rate for the detector, and it has not committed to a timeline for languages beyond English. Both gaps matter for anyone planning to cite a result in a moderation or compliance decision.
The business calculus
For platforms and publishers, the practical question is what a result authorises. A watermark hit can support a moderation action because it links the file to a specific generator. A miss cannot clear a file, and treating it as a clean bill of health would push risk onto the wrong side of the decision.
Trust and safety teams should log which detector version produced a result and preserve the original file, since the watermark's presence is itself evidence that routine processing can destroy. Teams handling user uploads also face a design choice: check files on ingest, when the watermark is most likely intact, or check them later, when context is richer but the signal may be gone.
The economics favour the check. Google carries the cost of the service and offers it free, so the marginal cost of verifying a file falls to the seconds a staff member spends uploading it. The follow-up, what a platform does once it has a result, carries the real cost.
Evidence handling adds a second layer of work. A watermark result is only as useful as the record that accompanies it, so teams that intend to rely on SynthID in a dispute need to store the file, the timestamp and the detector output together. Without that trail, a finding is hard to defend later.
The free, English-only launch also limits near-term reach. AI-generated misinformation spreads fastest in languages with thin fact-checking capacity, and those are precisely the markets the current release does not serve. Google has said nothing about when that changes.
Competitors face a narrower decision. Any large model developer that wants its output to be verifiable has to embed signals that Google's detector can read, which means coordinating with a rival's specification. Staying outside the coalition preserves independence and forfeits the verification benefit.
Why this matters
Google is pushing provenance from a research project into default infrastructure, and the public launch makes that ambition testable. The measure of success is not the launch itself but whether unmarked generators shrink and whether the detector's reach extends beyond English-speaking users.
For decision-makers, the immediate takeaway is to treat a SynthID result as one input among several: strong evidence when a watermark is found, no evidence when it is not. The next milestone to watch is Apple support going live, which would show how quickly the coalition can widen.
Sources
We're making it easier to identify AI-generated content globally.
Related Articles
- Google Launches Gemini 3.1 Flash TTS for AI Audio
- Nvidia Synthetic Video Detector: A Trust Play Built on Unverified Numbers
- Lyria 3.5 in Gemini Puts Full-Length Song Generation in Every User's Hands
✔Human Verified
Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.