SynthID detector: what it can check and what remains unconfirmed

What SynthID is and what the tool is designed to do

SynthID is a digital watermarking technology developed by Google to identify content generated by some of its artificial intelligence models. Google’s documentation says the watermark is embedded directly into the generated content. The feature discussed here does not judge every file or determine who created it: it looks for a particular SynthID watermark.

Google provides support documentation about verifying content, as well as general technical documentation on SynthID. Taken together, these pages explain the purpose of the watermark and confirm that verification covers images and videos. However, the available information is not enough to establish every detail of the reported public launch of a standalone detector, its general availability or the full process for accessing it.

That distinction matters. The existence of a watermarking technology, and Google’s documentation of a way to check for it, do not automatically answer questions such as who can use the tool, which files it accepts in practice or what information it returns in each case. Describing those conditions requires complete official instructions that apply to the specific tool.

Which formats it can check—and what remains unclear

Google’s support documentation covers checking images and videos. Those two formats therefore have documentary support as use cases for the detector. A support page’s title also mentions audio, but the verified information available here does not specify the conditions for audio analysis. It does not confirm which audio file types are accepted, which models or generation processes are covered, or whether the result works in the same way as it does for images and videos.

A general mention of formats should not be turned into a complete compatibility list. Compatibility may depend on the available tool, the type of content and how it was generated. Without sufficient official instructions for each case, it should not be assumed that every image, video or audio recording can be uploaded or verified with the same result.

In particular, it has not been confirmed here that the detector accepts every common file format, analyzes long files or identifies specific segments within any video or audio recording. These details should be checked in the official instructions provided when using the service. Their absence from the available information does not prove that a feature does not exist; it means that it should not be presented as verified.

What a positive result means

If a check identifies SynthID, the useful, limited conclusion is that the system detected a watermark associated with that technology in the content it analyzed, within the system’s capabilities. This may provide a clue about the content’s provenance in relation to Google tools that incorporate SynthID. It does not, on its own, prove who created the file, when they created it, what instructions they used or whether the content was published in its original form.

A watermark is not the same as a certificate of truthfulness. A marked image could contain false information, be shown out of context or depict an invented scene; detection does not verify the facts it portrays. Likewise, a watermark does not establish that someone manipulated the file with the intention to deceive. Those are separate questions that require other evidence.

Google’s general documentation describes SynthID as a watermark embedded in content generated by its AI models. That does not justify concluding that every creation from Google necessarily carries the watermark: the information available here does not establish that all models, products, versions or workflows are covered. Interpretation should stay within the result and the conditions stated by the tool.

Why a negative result does not rule out AI use

A negative result means only that the detector did not identify a SynthID watermark under the conditions of that check. It does not prove that the content was made entirely by people or that it was not created or altered with AI. The file may come from another provider, may have been generated without SynthID, or may contain a watermark that the system could not recognize in that analysis.

The tool should not be mistaken for a universal detector of synthetic images, videos or audio. It cannot conclude that content was not generated by AI just because it did not find the particular watermark it searches for. Nor does it replace a review of context: the file’s provenance, original publication, available metadata and comparison with other evidence may all be relevant, although none of these elements alone guarantees authenticity.

This limitation is not an anomaly; it follows from the test’s specific scope. A detector for a particular watermark answers a narrow question. Answering other questions requires additional signals and different methods. That is why the technical result and any editorial or investigative conclusions drawn from it should be communicated separately.

The number of marked items does not measure detector coverage

Some descriptions of SynthID cite a figure of more than 180 billion marked pieces of content. The official documentation available for this verification does not confirm the figure, specify the period it covers or clarify what counts as a piece of content. It should therefore not be presented here as an established fact.

Even if a figure of this kind were confirmed, it would not be equivalent to the number of files the detector can analyze or the share of all AI-generated content that carries a watermark. Interpreting it would require, at a minimum, a definition of the unit being counted, the time period, the products included and the counting method. It would not prove that the system recognizes every watermark or every modification to a file, either.

The scale of a watermarking technology and the coverage of a verification tool are different measures. The former concerns content to which a watermark has reportedly been applied; the latter concerns what the detector can process and recognize. Without comparable, documented information, one cannot be inferred from the other.

Limitations with other generators, edits and privacy

SynthID should not be confused with a universal standard shared by every provider. The documentation reviewed describes a Google technology; it does not confirm that the detector recognizes watermarks from other generators. If an image or recording carries a different watermark, that does not mean SynthID can identify it. And if a file has no compatible watermark, the result may be negative even if it was created using AI.

It is also important to distinguish watermark detection from the analysis of edits. The available information does not specify which changes preserve or remove detectability, or guarantee that a watermark can be recognized after every edit, compression or transformation. Without specific technical instructions that apply to the detector, claims of complete resilience or descriptions of successful cases would go beyond what has been verified.

Finally, the sources reviewed do not confirm how files submitted to the tool are handled: how long they are retained, who may access them or for what purposes they may be processed. Before uploading confidential, personal or restricted material, it is prudent to review the privacy and data-handling terms shown by the service. Users should not assume either that files are deleted immediately or that they are stored indefinitely.

What to check before using it

The confirmed information supports describing SynthID as a Google digital watermark, and Google’s support documentation allows us to discuss checks for images and videos. It is not enough, however, to confirm every aspect of the reported public launch, establish the conditions for audio, substantiate the figure for marked content or detail how submitted files are retained.

For anyone checking a file, the practical approach is to keep the conclusion narrow: identify the format being analyzed, record exactly what result the tool provides and do not extend that result to general authenticity or systems unrelated to SynthID. If sharing the result, explain what it checks and what it does not. For availability and privacy details, review the service’s current instructions before uploading content.

A detector’s usefulness depends as much on what it can recognize as on the limits of its answer. SynthID can provide a provenance clue when it detects its watermark. It is not a universal test for AI-generated content, nor does failure to detect the watermark certify that a file is authentic.

Keep exploring

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.