Google SynthID Detector Opens to the Public: What It Can Tell You

On October 7, 2026, Google opened the SynthID Detector to the general public. The SynthID Detector is a website at synthid.com where anyone can upload an image, video or audio file and check whether it carries a SynthID watermark. It is a useful step for AI-generated media detection, but it answers a narrower question than its name suggests: not "is this AI?" but "does this carry a watermark that Google or a partner embedded?"

This post covers what Google announced, how invisible watermarking differs from C2PA Content Credentials and from classifier-based detectors, where each approach breaks, and why regulation is pushing providers toward marking their outputs.

Google

OpenAI

NVIDIA

Apple

What Google announced

According to Google's DeepMind blog post by Pushmeet Kohli and TechCrunch's coverage, the key facts are:

• What it checks: whether an image, video or audio file was made with AI from Google or its partners, using SynthID's imperceptible watermarks. Neither the post nor TechCrunch describes text detection in this tool.
• Who can use it: anyone. An earlier version launched in May 2025 for journalists, media professionals and researchers. The public version is available globally, in English only.
• Partners: Google, OpenAI, NVIDIA and Kakao, with Apple listed as coming soon. TechCrunch also reports that OpenAI runs its own verification site.
• Formats: TechCrunch lists images (JPG, PNG, WEBP, HEIC and others), video (MP4, MOV, WEBM) and audio (WAV, MP3, FLAC, AAC and others). Google's post does not list formats.
• Scale: over 180 billion images and videos and 240,000 years of audio have been watermarked. Verification already built into Search, the Gemini app and Chrome handles over 1 million requests a day.

The Next Web adds, citing Engadget, that using the site requires signing in with a Google, OpenAI or Apple account. Google's post does not mention sign-in, rate limits or pricing, so treat that as reported rather than confirmed.

The growth figures come from comparing two Google posts. In its May 19, 2026 announcement, Google cited over 100 billion images and videos and 60,000 years of audio. By October the numbers were 180 billion and 240,000 years. Google does not explain the audio jump, and both are "over" totals.

How the three approaches work

Provenance tools fall into three families that answer different questions.

Invisible watermarking (SynthID). The generator embeds a signal in the content itself at creation time. Google's earlier DeepMind description says video watermarks are embedded in the frames and are imperceptible to viewers but detectable by a tool, while the text version adjusts token probabilities during generation. Because the signal lives in the pixels, samples or tokens, it can survive some edits that remove file metadata. A detector then reports whether the signal is present.

Content Credentials (C2PA). Here a file carries signed provenance data. Per the Content Authenticity Initiative, each asset is hashed and signed, creating a verifiable, tamper-evident record, so changes after signing can be detected. C2PA can record how media was made and modified whether or not AI was involved. In its May post, Google said Pixel 10 was the first smartphone to provide Content Credentials for images in its native camera app, so a credential can document that a camera captured something, which a watermark on AI output cannot do.

Classifier detectors. These are models trained to spot statistical traces of generation, with no cooperation from the generator. They can in principle flag content from any source, but they are probabilistic and are in a continuing contest with new generators.

Comparison of provenance approaches

The two cooperative systems are complementary rather than rival. OpenAI's verification page reportedly checks both SynthID and C2PA signals, per The Next Web, and Google's May post added C2PA verification to the Gemini app alongside SynthID. Google's October post, by contrast, does not mention C2PA.

Robustness and the limits

Google and the press are fairly direct about the limits:

• Only SynthID is detected. Other companies use different standards, so a clean result does not clear a file. The Next Web notes Microsoft and Meta run their own approaches.
• Edited versus generated is not distinguished. Sample results reportedly read "made or edited with Google AI" and warn the file may have changed since.
• Stripping is possible in some cases, per Engadget as relayed by The Next Web. Google's DeepMind page describes SynthID as not a silver bullet and not built to stop determined adversaries, and says confidence drops for text under thorough rewriting or translation.
• Cross-vendor reliability is uneven. TechCrunch cites a Reuters report that these tools often fail to identify content from their own makers' models, and The Next Web says Reuters found in July that Meta's checker missed Meta's own cropped fakes. That concerns Meta's tool, not SynthID, and Google has published no accuracy rates for the Detector.

The interoperability problem

A watermark is only as useful as the number of detectors that can read it. Today a verifier must support each mark separately: SynthID, Meta's system, Microsoft's, plus C2PA. Google's stated response is to widen participation, as OpenAI, NVIDIA and Kakao adopt SynthID, and in May it also named ElevenLabs and said it open-sourced its SynthID text technology. That is a de facto standard built by adoption rather than a neutral specification, and Google is separately on the C2PA steering committee. For users, the practical result is that a negative result from any single checker carries little weight.

There is also a coverage gap in the other direction. Open-weight image and video models run locally can be used without any mark, and no detector of this kind can see them.

Regulatory pull

Regulation explains much of the timing. Article 50(2) of the EU AI Act requires providers of AI systems that generate synthetic audio, image, video or text to mark outputs in a machine-readable, detectable format. The European Commission's code of practice page says the transparency obligations apply from August 2, 2026. As we covered in our OpenAI textGrain analysis, ActuIA reports systems already on the market before that date have until December 2, 2026 to comply with Article 50(2). Google's announcement does not cite the AI Act, so the link is context, not a stated motive. For the text side of the same story, see our post on Anthropic's invisible watermark.

What to watch

• Whether Apple's support ships and which of its content the Detector can read.
• Whether Google publishes accuracy, false-positive and robustness data for the Detector.
• Whether a public detector for text appears from Google or others. OpenAI's textGrain detector was not public at launch, per our earlier post.
• Whether C2PA and watermark checks converge in one verification flow.

For teams using several image and video models side by side, as on a multi-model platform like Metir, the takeaway is practical: outputs from different providers will carry different marks, so keep a record of which model made what.

Sources:

• Google expands SynthID Detector for AI content, Google DeepMind
• Google's new SynthID website can identify AI-generated media, TechCrunch
• Google opens SynthID Detector to all to spot AI images, video, audio, The Next Web
• Making it easier to understand how content was created and edited, Google
• Watermarking AI-generated text and video with SynthID, Google DeepMind
• How Content Credentials work, Content Authenticity Initiative
• EU Code of Practice on Transparency of AI-Generated Content, European Commission

Image credits

• Hero: Google campus buildings in Mountain View, California. Photo by Sebastian Bergmann, Wikimedia Commons, CC BY-SA 2.0.
• In-body: Google sign, Charleston Road, Mountain View. Photo by Dietmar Rabich, Wikimedia Commons, CC BY-SA 4.0.