SynthID Detector is public: what it checks for AI ad creative

Yes, anyone can now use SynthID Detector: Google opened it on October 7, 2026, at synthid.com, for images, video and audio, in English. It reads watermarks from Google and from partners Google names as OpenAI, NVIDIA and Kakao, with Apple listed as coming.

For a team shipping AI ad creative, the practical change is that a stranger, a client or a reviewer can run the same check you can. It tells you whether a supported watermark was found. It does not tell you whether a file is AI-free.

What Google says changed

Before this, the detector sat behind an early-tester waitlist for journalists and researchers. Google's announcement says it is now open to anyone, and that checks are also built into Search, the Gemini app and Chrome. Google says it has watermarked over 180 billion images and videos and 240,000 years of audio.

Press coverage adds two access details that the Google post text I fetched does not state, so treat them as reported: you sign in with a Google, OpenAI or Apple account, and each person gets roughly 10 checks a day.

What a result does and does not prove

Google's own wording is that this is not a general AI detector and that it can only detect media from companies that adopted SynthID. Reports of the launch add that a hit reads as made or edited with Google AI, so a light touch-up and a fully generated image get the same label.

A miss is weaker still. Tools from non-partner companies, older systems, and files whose signal was degraded by edits can all come back empty. So a clean result is not evidence of a human-made file, and a hit is not proof of how much AI was used.

• Hit: a supported watermark was found. Disclose accordingly on platforms that ask.
• Miss: no supported watermark found. Say nothing about the file being real.
• Upload the highest-quality version you have; reports of the launch say that gives the most accurate result.

What this means for files that came out of Sume

The Sume docs pages I read for video and trim do not say that Sume adds its own watermark or Content Credentials to output, so do not promise one to a client. Whether a clip carries a partner's watermark depends on the model that generated it, and that is a question for the model vendor's own page.

Edits matter too. Sume's video trim doc says the default exact precision is a frame-accurate re-encode (libx264, yuv420p), while keyframe precision is a stream copy. Either way the output is a new MP4, not your original file, so check the file you will actually deliver, not the one you generated.

Run the check on the final export. If you trimmed, captioned or cropped after generation, a hit on the raw clip does not guarantee a hit on the delivered one.

A short routine before you publish

Keep the job record for each clip next to the delivered file, check the final export, and write down the result with the date, since the detector and its partner list are changing quickly. If a platform asks for an AI label, set it from what you know about how the file was made, not from what the detector returned.

• Check the delivered file, not the raw generation.
• Log date, file name and result (hit or miss).
• Label from your production record, never from a miss.