Google opens SynthID Detector to everyone: one site checks AI media from Google, OpenAI, NVIDIA and Kakao

Is that viral photo real, or did someone type it into an image generator five minutes ago? Until now, answering that question for content made with mainstream AI tools was surprisingly awkward. On 7 October 2026, Google made it a lot simpler. Its SynthID Detector is now open to everyone, worldwide and in English, on a dedicated website at synthid.com. Upload an image, a video or an audio file, and the site tells you whether it carries a SynthID watermark, the invisible signal that Google and a growing list of partners hide inside AI generated media.

The bigger change is not the website itself, but what it can read. For the first time, a single public checker looks for the watermarks of several companies at once: Google, OpenAI, NVIDIA and Kakao, with Apple announced as "soon". That turns SynthID from a Google feature into something closer to a shared industry tool. Here is how it works, what it can and cannot do, and why a free watermark checker is both very useful and easy to misread.

What Google announced

The announcement came in a post on Google's blog by Pushmeet Kohli, Vice President for Science and Strategic Initiatives at Google DeepMind. Google introduced SynthID in 2023 as a way to mark AI images with a signal that people cannot see but software can detect. Since then the system has grown to cover video, audio and text. According to Google, it has so far watermarked more than 180 billion images and videos, plus 240,000 years of audio. Those are Google's own figures, and they show how much synthetic media already carries the mark.

Last year, Google showed an early version of the SynthID Detector to journalists, researchers and other media professionals. That version was invite only. Everyone else had one practical option: upload a file to the Gemini app and ask it. Ars Technica points out that this was clunky, because a chatbot answer is not always the clear yes or no you want when you are checking a suspicious picture.

Now the Detector is a normal website. You sign in with a Google, OpenAI or Apple account, drop in a file, and get a result. Google says the new site joins the verification features it already built into Search, the Gemini app and Chrome, which together handle more than one million verification requests per day.

How SynthID works

SynthID is a watermark, not a guess. That difference is the most important thing to understand about this tool. A partner model adds the mark at the moment it creates or edits a file. For images and video, the signal is woven into the pixels. For audio, it sits in the frequency components of the sound. Unlike a visible logo or a metadata tag, it is designed to survive the ordinary life of a file on the internet: cropping, filters, color changes, compression and reposting.

When you upload something, the Detector looks for that intentional signal. According to the FAQ on synthid.com, the answer comes in two forms. Watermark detected means the file was likely generated or edited with a SynthID enabled model. Watermark not detected means no confident signal was found. For videos and audio files, the site can also point to the segments that carry the watermark, which matters for clips that mix real footage with AI generated parts.

The public site gives less detail than the internal tool that testers used. Ars Technica reports that the earlier version highlighted the regions of an image where it found SynthID pixels. The new site only tells you whether SynthID was used to produce or edit the media. Engadget adds that it does not say whether a file was fully created by AI or only touched up, so you will not learn which parts are synthetic in a still image.

One checker, several companies

This is the part of the news that makes the biggest practical difference. Ars Technica explains that Google's previous detector only looked for Google's own version of SynthID. Even though images from ChatGPT carry SynthID, Google's checker would not flag them. Android Authority describes the same split from the other side: the Gemini app only detects Google watermarks, and OpenAI's own verification site only detects watermarks from OpenAI tools.

The new site reads the distinct watermarks of all current partners. On Google's side that covers its Nano Banana image models, Veo for video, Lyria for music and audio, and products such as Gemini, Flow, ProducerAI and Vids, according to TechCrunch. OpenAI, NVIDIA and Kakao already use SynthID. Apple is coming next, which Engadget says should make the site one of the first tools able to recognize images made or edited with Apple Intelligence features such as Image Playground. Engadget also notes that ElevenLabs had previously committed to support the watermark.

The limits you need to know

Google is unusually direct about what the tool is not. The site itself says it is "not a general AI detector". It can only find media from companies that have adopted SynthID. That leaves large gaps.

• Other watermark systems. Meta and Microsoft use their own watermarking and verification approaches. Engadget mentions that Meta has been testing its own web tool, which recognizes images from its latest Muse Image models, but not those from other companies or even its own older models.
• Open weight models. Anyone can run an open image or video model on a home computer. Ars Technica notes that this kind of output often has no watermark and not even basic metadata labels. SynthID Detector will say "not detected" for all of it.
• Removal attempts. The FAQ admits that the watermark is not infallible. Someone who tries many transformations may find one that breaks the signal. Heavy compression and editing can weaken it too, which is why Google recommends uploading the highest quality version of a file you can find.
• Rare false alarms. The FAQ also says that in very rare cases SynthID can trigger on content that was never watermarked, and it recommends collecting several pieces of evidence before deciding.

The daily limit is another constraint. Google gives each user "a daily usage quota of approximately 10 image, video, and audio checks", according to Ars Technica. Google engineers told the publication that the cap exists to stop people from using the checker to build tools that strip the watermark. The vague "approximately" probably means the system can lock you out earlier if you upload many very similar files, which looks like someone tuning a bypass. The login requirement has the same purpose, but Android Authority points out the obvious downside: if you do not want an account with Google, OpenAI or Apple, you cannot use it. A Google representative confirmed to the site that there is no other way in.

What you can upload

The site accepts a wide range of everyday files. According to TechCrunch, that means 11 image formats including JPG, PNG, WEBP, HEIC and GIF, three video formats (MP4, MOV and WEBM) and six audio formats, from WAV and MP3 to FLAC and M4A. In practice, almost anything you save from a messenger, a social network or a phone camera roll should work.

A practical tip follows from the limits above. A screenshot of a screenshot that has been shared through three apps is the worst case for any watermark. If you can, track down the original post or the earliest copy of a file before you test it. And if the answer is "not detected", treat that as "no SynthID found", not as proof that a picture is real.

How we got here

SynthID did not appear overnight. Google DeepMind launched it in August 2023 for images from its Imagen model. In October 2024 it open sourced SynthID Text, the version that marks text written by language models. In May 2025, at Google I/O, it showed the Detector portal to early testers. Since then, OpenAI, NVIDIA and Kakao have joined as partners, and the public launch on 7 October 2026 is the step that finally puts the checker in everyone's hands.

The timing also fits a wider trend. Regulators increasingly expect AI providers to mark synthetic media in a machine readable way, and the EU AI Act includes transparency duties for AI generated content. We recently covered how OpenAI started adding invisible watermarks to ChatGPT and Codex text in the EU. A shared checker for image, video and audio marks is the natural next piece of that puzzle.

Why it matters

Deepfakes are no longer a niche problem. Realistic AI images and voice clips show up in elections, scams, celebrity hoaxes and family chat groups. Most people do not have forensic tools, and general "AI detectors" that guess from visual artifacts are often unreliable. A free website that can confirm a hidden, intentional signal is a real improvement for anyone who wants to check something before sharing it, from journalists and teachers to the person who receives a suspicious voice message.

The multi company support is what gives the site weight. When the biggest image and video generators share one watermark and one checker, a single test covers a large share of the AI media people actually see online. Every new partner makes the "detected" answer more useful.

At the same time, the launch exposes the core weakness of watermarking. It only works for honest players who agree to mark their output. Bad actors can use tools that never add a mark, or try to strip it. That is why Ars Technica argues the long term answer may be the opposite approach: labeling real photos, videos and recordings at the moment they are captured, with cryptographic standards such as C2PA. Google's Pixel phones already do this, and Google's Circle to Search can check both SynthID and C2PA credentials, according to Engadget. Watermarks for AI content and signatures for real content work best together.

Dany's take

I like this launch, and I tried to think about who it really helps. A journalist with a deadline or a parent who gets a strange video of their kid's school can now check it in seconds, without a special tool and without asking a chatbot to interpret the result. That is real progress, and Google deserves credit for opening the checker to its competitors' watermarks too.

But I am a bit worried about how people will read the green and red answers. "Watermark detected" is strong evidence. "Not detected" is almost no evidence at all, because so much AI content comes from tools that never add SynthID in the first place. My fear is that someone uploads a fake made with an open model, gets "not detected" and shares it as proof that the image is real. Google writes this caveat on the site, but caveats rarely go viral.

I also do not love the account requirement. I understand the reason, because an unlimited checker would be a perfect training tool for people who want to remove watermarks. Still, a tool meant to protect the public should not require a Big Tech account. My wish list for the next update is short: an option without login for basic checks, Meta and Microsoft joining the same system or at least a shared checker, and more devices signing real photos with C2PA. Until then, use synthid.com as one useful clue, not as a final verdict.

Sources

• Google blog: Google expands SynthID Detector for AI content (Pushmeet Kohli, 7 October 2026)
• SynthID Detector website and FAQ
• Ars Technica: Google rolls out improved SynthID AI content detector, now available globally (Ryan Whitwam, 7 October 2026)
• TechCrunch: Google's new SynthID website can identify AI generated media (Ivan Mehta, 7 October 2026)
• Engadget: Google's AI detection website is now available (Karissa Bell, 7 October 2026)
• Android Authority: Google's cross model AI detection tool is finally available for all (7 October 2026)
• The Verge: Google's AI content detector is rolling out globally (Emma Roth, 7 October 2026)