Google has opened its SynthID verification site to the public. As TechCrunch reported on October 7, 2026, anyone can now upload an image, video or audio file at synthid.com and see whether it carries Google's invisible AI watermark. Since Google I/O 2025 the detector had been a limited-access portal for journalists, media professionals and researchers. Now it is a front door for ordinary people who just received a suspicious clip in a group chat.
The headline is useful, but the caveat is the more important half. A positive result is strong evidence that a file came out of a SynthID-enabled model. A negative result proves almost nothing. This post explains what the site does, what the coverage confirms and does not confirm, and how to fold a SynthID check into a sensible verification routine. It builds on our earlier explainers on how AI watermarking works and what C2PA Content Credentials are.
TL;DR: what changed on October 7
| Question | Answer |
|---|---|
| What launched? | A public SynthID website for checking media, per TechCrunch |
| Who could use SynthID Detector before? | Journalists, media professionals and researchers via a waitlist |
| Media types | Images, video and audio (the report lists no text support on the site) |
| Image formats | JPG, JPEG, PNG, BMP, WEBP, AVIF, HEIC, HEIF, TIFF, TIF, GIF |
| Video formats | MP4, MOV, WEBM |
| Audio formats | WAV, MP3, OGG, FLAC, AAC, M4A |
| Scale reported | About 1 million verification requests per day |
| Other surfaces | Verification is also built into the Gemini app and Chrome |
| Who else uses SynthID? | OpenAI, NVIDIA and Kakao, according to the report |
| Biggest caveat | The tools are described as not infallible and can miss content, even from Google's own models |
A short history: from waitlist portal to public site
Google DeepMind introduced SynthID in 2023 as a way to embed an imperceptible signal in AI-generated images. The idea expanded to audio, video and text. On May 20, 2025, Google announced the SynthID Detector portal, a single place to upload content and see whether a watermark is present, including highlighting which portions are most likely watermarked. At that time Google said over 10 billion pieces of content had been watermarked with SynthID, and it opened a waitlist for early testers.
In the same announcement Google named NVIDIA as a partner that would watermark videos generated by its Cosmos models, and GetReal Security as a verification partner. Pushmeet Kohli of Google DeepMind described the design goal in that post: the watermark "remains detectable even when the content is shared or undergoes a range of transformations."
Since then SynthID has spread sideways. Google applied the same family of ideas to biology in September, which we covered in SynthID Bio: DeepMind can watermark AI-designed proteins. New Google generators ship with the mark by default, such as the music model covered in Lyria 3.5 in Gemini and the image model in Nano Banana 2.1. The October 7 launch is the consumer-facing end of that pipeline: the thing a non-specialist can actually open.
What the site does, step by step
Based on the launch report, the workflow is deliberately simple. You open the site, choose a file, and the detector scans it for a SynthID signal. The answer tells you whether a watermark was found. Because the format list is broad, including HEIC and HEIF from phones and common audio containers like M4A, you can usually test a file straight from a messaging app download without converting it first.
Two practical points follow from how watermarks work:
- Test the original file whenever possible. Each re-encode, crop, filter or screenshot degrades the signal a little. A watermark that survives one JPEG recompression may not survive a screenshot of a screenshot.
- Test several copies if you have them. The same clip posted on three platforms has been compressed three different ways. One copy may preserve the mark when another does not.
What the coverage confirms, and what it does not
TechCrunch reports the supported formats, the 1 million daily requests figure, the Gemini and Chrome integrations, and the claim that OpenAI, NVIDIA and Kakao support SynthID. It also quotes the limitation clearly: the tools are not infallible and often fail to identify content, including media created by Google's own models.
What the report does not give us is a false-positive rate, a false-negative rate, upload size limits, rate limits, pricing, a privacy statement about uploaded files, or a list of which third-party generators actually embed the mark. If you are building a newsroom or trust-and-safety workflow around this site, those gaps matter. Treat the site as an additional signal, not a verdict engine, until Google publishes accuracy figures.
How SynthID compares with C2PA and text watermarks
People often ask whether SynthID replaces Content Credentials. It does not. They solve different halves of the problem.
| Property | SynthID watermark | C2PA Content Credentials |
|---|---|---|
| Where the signal lives | Inside pixels, audio samples or token choices | In signed metadata attached to the file |
| Survives a screenshot? | Often partly, not guaranteed | Usually stripped |
| Can be verified by anyone? | Only through the vendor's detector | Anyone with a C2PA reader |
| What it proves | This came from a model that applies the mark | Who signed it and what edits were declared |
| Works on text? | Yes, as a statistical bias in word choice | Weakly, since text has no container |
That is why labs increasingly ship both. When we examined Anthropic's Claude watermarks and C2PA, the reasoning was the same: metadata is easy to read but easy to lose, while an embedded mark is hard to see but needs the right detector. OpenAI's approach to text, described in our textGrain explainer, is another vendor-specific mark with its own detector access rules. The practical consequence is fragmentation: a clean result from one vendor's detector says nothing about another vendor's output.
Why a negative result is weak evidence
Imagine you upload a viral video and the site finds no watermark. There are at least five innocent explanations that leave the video fully synthetic:
- The generator never applied SynthID.
- The clip was produced by an open-weights model run locally, where no one is obliged to mark output.
- The file was re-encoded, cropped or laundered through an editing tool until the signal faded. Our guide to AI watermark removal covers why that is possible and where the law stands.
- The detector missed a mark that was present, which Google concedes can happen.
- The generator is from a vendor that uses a different watermark entirely.
For that reason the question "will every model eventually mark its output?" matters so much, and we looked at it in will all AI models watermark output. Until the answer is yes for open-weights models, absence of a mark will stay unconvincing.
A practical verification routine
If you review media for a living, or just want to avoid forwarding fakes, treat SynthID as one step in a layered routine:
- Check provenance metadata first. Look for Content Credentials with a C2PA reader. If present and valid, read what edits were declared.
- Run the SynthID site. A hit is meaningful. Record the file hash and the date, since detectors improve and results can change.
- Trace the source. Find the earliest upload, the account that posted it, and whether anyone reputable corroborates the event.
- Look for physical and contextual inconsistencies. Shadows, hands, lip sync, audio room tone, and impossible details.
- Do not conclude real from a clean scan. Say "no SynthID watermark detected," never "verified authentic."
For teachers and students, our piece on what watermarking means in classrooms applies the same caution: a detector is a conversation starter, not a verdict.
What it means for builders and publishers
If you ship a product that generates media with Google models, the mark is already applied for you. Your obligations are mostly about disclosure and not stripping it. If you ship a product that ingests user uploads, you now have a public reference detector that your users may point at your output, so expect questions when a result comes back positive or negative. It is worth documenting in your help center which generators you use and whether they mark output.
If you run moderation or a newsroom desk, the signal to watch is not a single result but the cumulative effect: more vendors adopting a common mark makes a shared detector more useful, and Google reportedly now counts OpenAI among them for some products. That would be a notable shift in a space where every lab historically built its own island. Whether other companies will accept a Google-hosted detector as neutral ground is an open question, and one explainx.ai will keep tracking.
There is also an incentive story. Watermark detection at scale, 1 million requests per day by the reported count, is a data advantage and a distribution advantage. We explored the business side in are AI watermarks monetisable. The more people rely on a vendor detector for trust decisions, the more that vendor shapes what counts as verifiably synthetic.
Open questions to watch
- Accuracy numbers. Will Google publish measured detection rates across common edits like cropping, compression and re-recording?
- Privacy. What happens to uploaded files, and are they used to improve the detector?
- Third-party coverage. Which of OpenAI's products actually embed SynthID, and in which media types?
- Adversarial pressure. Removal tools are already a cottage industry. Expect studies on how quickly the new public site is attacked.
- Regulation. The EU's transparency rules push labs toward marking output, and a public detector makes compliance visible.
Bottom line
The SynthID website turns a specialist tool into a public utility, and that is a real improvement in the information environment. Use it, but use it honestly: a hit tells you something, a miss tells you almost nothing, and no detector replaces sourcing and judgment. We will update this post when Google publishes accuracy data or terms.
Related reading
- How does AI watermarking work? Text explained
- What is C2PA and Content Credentials?
- SynthID Bio: DeepMind watermarks AI proteins
- OpenAI textGrain text watermark
- Anthropic Claude invisible watermarks and C2PA
- Will all AI models watermark output?
- Official: Google's SynthID Detector announcement and synthid.com
Details reflect reporting as of October 7, 2026; supported formats, limits and partner lists may change.