Google opened SynthID Detector to public use in an October 7, 2026 announcement, offering worldwide access in English to check images, video and audio for supported AI watermarks. For anyone deciding whether to share a striking picture or reuse a recording, it provides an extra clue about how the file was made.
The important limit: finding a watermark is evidence of AI involvement, but finding none does not prove human authorship. Neither result settles whether the event, caption or spoken claim is true.
1. Keep the source and choose what to upload
Save the original post’s link, its caption and the best available copy of the file. Keeping those together gives you something to check beyond the detector’s answer. The Detector’s upload guidance recommends using the highest-quality file available.
Before uploading private or work material, review the service’s data handling. Google’s portal privacy notice says it deletes the original media after returning results, but keeps an account-linked digital signature of it for 24 hours. File deletion does not mean no associated data is retained.
2. Choose the right watermark check
SynthID Detector and Gemini offer two different watermark checks.
SynthID Detector: DeepMind lists coverage for supported media from Google, OpenAI, NVIDIA and Kakao. Use this route for the broader provider check—not as a detector for every AI product or every older file.
Gemini: Its SynthID check recognizes only Google AI. Gemini can also inspect compatible Content Credentials, a separate source of information about a file’s history.
Apple coverage remains unclear: Google’s announcement includes Apple, while DeepMind’s product page says it is coming. Do not assume Apple-generated media is already covered.
SynthID is an imperceptible signal embedded in generated content, not a visible badge or an ordinary metadata label. The public Detector lists images, video and audio; do not assume that SynthID’s separate text-watermarking technology makes this a text-checking service.
3. Submit the file
In the standalone Detector: Visit SynthID Detector, review its notice, follow the sign-in flow and use Select file. Read the result for the uploaded copy. Follow the portal’s own upload and usage limits; Gemini’s numbers below are not portal limits.
In Gemini: Sign in with a personal or qualified Workspace account, using the current app or Gemini on the web. Choose Add files and upload one file. Ask it to check whether Google AI created or edited the media. Google documents the upload-and-question approach and the current verification requirements.
For Gemini, files must be 100 MB or less, videos under 90 seconds and audio under one hour. Its approximate rolling 24-hour quota is 10 checks per media type, with five minutes of video and three hours of audio in total. These are usage limits, not guaranteed capacity.
4. Read the result without turning it into a real-or-fake verdict
Use four outcomes in your own notes. These are practical categories, not a promise that every interface uses identical labels:
Outcome | Meaning and next step |
|---|---|
Watermark detected | Treat this as evidence of a recognized watermark. Check the provider and any marked sections before describing what was AI-generated. The portal’s FAQ also acknowledges rare false positives. |
No watermark detected | AI involvement remains possible: unsupported providers, files made before adoption and heavy modification can escape detection. Seek other evidence; do not label the file human-made. |
Uncertain | Leave the result unresolved. Gemini documents uncertainty with simple imagery or very small edits; an inconclusive answer is not a clean bill of authenticity. |
Check unavailable | An upload failure, access problem or exhausted quota means no usable check was completed. Record that separately from a negative result. |
For video, pay attention to the audio and visuals separately. Google’s video-verification explanation describes checks across both tracks and responses identifying time ranges. An audio watermark is not a verdict on the pictures, and a marked section is not proof that the entire recording was generated.
The practical implication of the different tool scopes is that their answers need not match. A negative Gemini watermark result does not rule out a partner watermark that the standalone Detector may recognize. Keep the tool’s name with the result.
5. Check the claim attached to the media
If the file supports an important claim, find its earliest available source, check the stated date and place, and look for independent corroboration. A hypothetical example shows why: a genuine camera photograph can be paired with a false caption. No AI watermark is needed to mislead.
Content Credentials can add signed information about creation and editing. They describe recorded provenance—the file’s history—not whether an attached claim is true. Gemini supports compatible credentials on the web and Android, with iOS listed as upcoming in its current help page. Missing credentials alone do not establish manipulation.
For a small team, a useful record is the source link, exact file checked, tool, time, result and any marked sections. If the underlying claim still lacks support, leave it unverified rather than presenting the detector’s answer as permission to share.
Teams building audio workflows can also read RohitAI’s guide to supported OpenAI audio watermarking. That developer-focused workflow is separate from these manual media checks.