Google has made its AI watermark detector SynthID publicly available, allowing anyone to check whether images, videos, or audio were created by Google's AI or partners like OpenAI and Nvidia. More than 180 billion pieces of content already carry the watermark, but the tool only identifies SynthID-tagged content, not AI-generated media in general.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Google has opened the SynthID detector to everyone, but its capability boundary is as clear as its coverage: it only recognizes watermarks embedded by its own and partners' systems, not a universal AI detector.
Google has opened the SynthID detector to the public, allowing anyone to check whether an image, a video, or an audio clip was generated by its AI or partners. The tool supports four formats: JPG, PNG, MP4, and MP3, and can identify invisible watermarks embedded in content from models such as Gemini, Veo, and Lyria.
But the key qualifier in this opening is "SynthID": it is not a universal AI detector and can only flag content carrying a SynthID watermark. In other words, it answers "does this content carry our watermark," not "was this content generated by AI."
Placing SynthID's capability boundary back into the industry context reveals it is not an isolated case. Other detection systems have the same limitation, typically only able to verify their own provider's watermarks. This means the current watermark detection ecosystem is fragmented: each vendor can confirm its own marks but struggles to confirm others'.
From this, it can be inferred that the detector's coverage is limited to models and platforms that adopt this watermark—it cannot determine general AI-generated media. For users, a detection result of "no watermark found" does not equal "content not AI-generated," and this distinction is the premise for understanding the tool's value.
SynthID's coverage is not limited to Google's own models. OpenAI, Nvidia, and Kakao also use SynthID in media content. This explains why the detector can recognize "partner-generated" content.
But in text detection, the path forks: OpenAI developed its own system. Text is not within SynthID's coverage, which is also consistent with the industry pattern of "only recognizing one's own watermark."
Detection capability does not exist only on a standalone page. SynthID detection is already built into Google Search, the Gemini app, and Chrome. According to Google, Apple is expected to join soon—this is Google's unilateral expectation, and the timing and specific form of joining have not been confirmed.
Google says its platforms receive one million verification requests per day and claims that over 180 billion images and videos have been watermarked to date. These two figures sketch a sense of scale, but both are Google's unilateral claims and lack third-party verification.
Also not explained by the source are: the robustness of SynthID watermarks against tampering or removal, and the detector's actual accuracy and false positive rate after public opening. These gaps mean that scale figures cannot be directly equated with detection reliability.
The facts in this article are based on THE DECODER's report on the opening of Google's SynthID detector. The detector's opening, supported formats, covered models, not being a universal detector, partners using SynthID, OpenAI's self-developed text system, and integration into Search/Gemini/Chrome are confirmed facts; Apple's expected joining, one million daily verification requests, and 180 billion watermarked items are all Google's unilateral claims without third-party verification; the judgment that coverage is limited is an inference based on the source.
The significance of opening the SynthID detector lies in handing the judgment of "does it carry our watermark" to the public, rather than providing a universal key for AI content recognition. Its value is as clear as its boundary: it can only confirm marks within the SynthID ecosystem, and non-detection does not equal non-AI generation.
Primary report
Primary source