How Google’s SynthID detector checks watermarks, not AI authorship

Engadget reports that the site can flag watermarked content from AI tools associated with Google, OpenAI, NVIDIA and Kakao. Those provider names describe the coverage reported by the outlet; they are not the result of an independent test measuring how reliably the website detects each provider’s output.

The report says the service is free, but users must sign in with an account from Google, OpenAI or Apple. Having Apple as a sign-in option does not, on its own, mean the detector can identify content made with Apple’s AI features.

The public access follows an earlier, more limited offering to journalists and other media professionals, Engadget reported. That change affects who can try the tool, not what a result establishes about a file’s origin.

Engadget’s description of the public site lists images, video and audio. It does not confirm that people can submit text, so the site should not be treated as an AI-writing detector based on the information available about this launch.

What does a SynthID result tell you?

A positive result indicates that the checker found a SynthID watermark in the submitted content. It does not, by itself, show that the entire file was generated by AI or explain how much AI contributed to it.

Engadget says the website does not distinguish content created entirely with AI from content edited with AI. Nor does it identify which parts of an image, video or recording were generated or changed. The result is a signal about a watermark, not a complete record of the file’s production.

That distinction matters for interpreting a marked file. A user might learn that the detector found SynthID, but the site does not tell them whether AI created the whole item or was used to modify it. The result alone cannot answer that broader question.

A result with no detected watermark is limited, too. Engadget reports that not all AI systems use SynthID and that watermarks may be removed in some situations. AI-generated or AI-edited content could therefore receive no positive result.

Is Apple Intelligence detection available?

Engadget describes detection for images made or edited with Apple Intelligence as forthcoming. The same report also characterizes the website as among the first tools able to identify images from Apple features, leaving the present status unclear. It does not confirm that Apple Intelligence detection is currently available or give a date for its arrival.

The distinction between account access and detection support is important: signing in with an Apple account is one of the listed ways to use the website, but that does not confirm that the tool can already detect Apple Intelligence watermarks.

How reliable is the public detector?

The launch report does not establish how accurately the website works. No independent evaluation of the public site’s accuracy or coverage was found in the reporting reviewed. The available information does not provide measured rates for missed watermarks or incorrect detections.

The list of provider-associated tools is not a substitute for those measurements. It does not establish how the detector performs across particular models, versions, or files that have been edited or combined with other material. Users also cannot infer from the launch alone how often a mark will be recognized in the files they encounter.

For now, the supported conclusion is narrow: SynthID Detector checks supported media for a watermark it can recognize. A positive result does not explain how much of a file was made or edited with AI, and a negative result does not prove that a person made it. An independent evaluation of accuracy and coverage would help show how much confidence users should place in either result.