ai-toolshow-totutorialsgooglesecurity

How to Check AI Files with the Public SynthID Detector

Upload an image, video, or audio file to Google’s public SynthID Detector. Learn what a watermark hit means and which partners it covers.

October 7, 2026 Yum News

A sharp photo in a group chat is no longer proof a camera took it. On 7 October 2026, Google opened the SynthID Detector to everyone, in English, worldwide. You upload an image, video, or audio file at synthid.com and the portal looks for a SynthID watermark written by Google or a partner.

This is not the Gemini chat check. That flow still works, and we covered it in How to Check Images for SynthID Watermarks in Gemini. The public detector is a separate site. Last year it was limited to journalists and media professionals. Pushmeet Kohli, VP of Science and Strategic Initiatives at Google DeepMind, said the broader version is available starting today.

What the detector can and cannot see

SynthID is a watermark Google DeepMind launched in 2023. It is woven into pixels, video frames, or audio frequencies when participating tools generate or edit media. Google says the mark is imperceptible and is meant to survive ordinary sharing, such as filters, colour changes, and lossy compression.

The public portal is explicit about its scope. It is not a general AI detector. It can only report on media from companies that adopted SynthID. Google’s 7 October post names Google, OpenAI, NVIDIA, and Kakao, and says Apple support is coming. The detector page says the list grows as more organisations join.

Google also published scale figures with the launch. It says SynthID has watermarked more than 180 billion images and videos, plus 240,000 years of audio. Built-in checks in Search, the Gemini app, and Chrome now handle more than 1 million requests a day. Those numbers come from Google, not from an independent audit.

A miss does not prove a file is a photograph or a human recording. The portal’s own FAQ says a “not detected” result means no confident SynthID signal was found. The file may come from a model that does not write SynthID, or heavy edits may have degraded the mark. In rare cases the detector can also flag non-watermarked content.

Open the portal and upload the best file you have

You do not need a newsroom account. Open synthid.com in a browser. The page is in English.

Choose the media type that matches the file: image, video, or audio. The detector analyses those three types.
Drag the file onto the upload area, or use Select file.
Prefer the original download over a screenshot, a messaging-app preview, or a re-exported copy. The site asks for the highest-quality file you have, because compression and editing weaken the signal.
Wait for the scan. Do not refresh mid-upload if the file is large.
Read the result before you caption or republish the file.

The May 2025 journalist portal also accepted text. The public page launched on 7 October describes image, video, and audio checks. Use Gemini if you need a text check, and treat that as a different test.

Person working on a laptop while reviewing digital files

Read the two results the portal returns

The FAQ on the detector site defines the outcomes in plain language.

Watermark detected. A SynthID mark is present. The file was likely edited or generated by a SynthID-enabled model. The older journalist portal highlighted portions of an image and segments of audio or video. If the public result shows a region or a time range, record that detail. A partial hit can mean a real frame with a generated object composited in.

Watermark not detected. No confident signal was found, so the file was not likely edited or generated by a SynthID-enabled model. The same FAQ says this does not rule out AI generation or manipulation.

Write the claim to match the result. “SynthID watermark detected on 7 October 2026” is a fair note. “Verified real” is not, even after a clean scan.

Run a known-file test first

A one-time control makes the wording easier to trust.

Generate an image in the Gemini app with a plain prompt, such as “a blue enamel mug on a steel counter, photo.” Download the file Google returns.
Upload that download to the SynthID Detector. Expect a watermark hit if the file still carries the mark Google writes into Gemini image output.
Upload a photo you took with your phone, exported from the camera roll without a generative edit. Expect no confident watermark.
Save both results, including the date. Wording on the page can change. The useful record is which outcome you got.

If the generated file misses, retry with the original download. A screenshot of the chat often drops the mark. Google says the watermark is robust to common edits, and also says it is not infallible if someone stacks many transformations.

Partners, Chrome, and Gemini are separate checks

A file from ChatGPT, an NVIDIA tool, or Kakao may carry SynthID if that product writes the mark. Google says Apple support is coming, not that it is live today. A partner miss can still mean the product has not enabled the watermark on that output yet.

Search, the Gemini app, and Chrome already run SynthID checks. DeepMind’s product page says you can upload an image, video, or audio clip and ask whether Google AI created or altered it. That answer is about Google’s mark inside those products. The public detector is the place to check partner media in one upload, without opening a chat.

Content Credentials (C2PA) are a different label. They record a story of the file, such as camera capture or an AI edit, when the tool writes them. A SynthID scan does not replace a credential, and a credential does not replace a watermark. Use both when the file offers both.

Protein watermarks are unrelated to this portal. SynthID Bio, announced on 30 September 2026, marks AI-designed protein sequences. It does not scan photos.

Abstract view of Earth from space, used as a stand-in for global media checks

Limits and rules worth stating

Match the label to the tool if you publish a finding.

Name the portal, the date, and the exact outcome. Do not upgrade a miss into proof the file is human-made.
Upload the fullest file you have, then crop for display. A cropped preview is a weaker sample.
The detector site says heavy modification can degrade the signal. It also says false triggers on unmarked content are very rare, not impossible.
Google’s additional terms ban using the service to attack or bypass watermark detection, and they ban scraping the page as a bulk API. Check files you have a reason to review. Do not script the portal.
The page is English-only at launch. A non-English interface is not documented in the 7 October post.

If you already check files inside Gemini, keep that habit for Google-made media and use the portal when the source might be a partner. The two results can disagree if one tool saw a degraded copy. Keep the original and note which check you ran.

Conclusion

The public SynthID Detector is a file upload, not a verdict on every picture on the internet. Open synthid.com, send the highest-quality image, video, or audio file you have, and record whether a watermark was detected. A hit means a participating model likely generated or edited that media. A miss means the portal did not find a confident SynthID signal.

Run the mug test once so you know what a positive looks like on your account. After that, pair the portal with the Gemini upload flow when the file might have come from Google, and say only what the scan actually returned.