The logos of Google, OpenAI, Nvidia and Kakao, the companies whose AI media the SynthID Detector can check. Illustrative. Image: Google, OpenAI, Nvidia and Kakao logos via Simple Icons (CC0)
Google has opened its SynthID Detector to everyone: a free website where you can upload an image, a video or an audio clip and find out whether it carries the invisible watermark that Google, OpenAI, Nvidia and Kakao put into media made by their AI.
Until now the tool was limited to selected journalists, media professionals and researchers, who got to test it after Google showed it off at its I/O conference last year, TechCrunch reported.
How it works
SynthID is Google DeepMind’s watermarking system. Instead of a visible logo or a metadata tag that can be stripped out, it hides a signal in the pixels of an image or video and in the frequency components of audio. Google says the mark is invisible to people but survives filters, colour changes and lossy compression.
The detector supports three kinds of media:
- Images: it says whether a watermark is present.
- Video: it points to the segments that carry the watermark.
- Audio: it does the same for clips of speech or music.
It accepts most common formats, from JPG, PNG, WEBP and HEIC images to MP4 and MOV video and MP3, WAV and FLAC audio. Google’s own Nano Banana, Veo and Lyria models watermark everything they make, as do apps built on them such as Gemini, Flow and Vids. SynthID checks are already built into the Gemini app and Chrome, and Google says people make about a million verification requests a day, according to TechCrunch.
“This is not a general AI detector”
The site is blunt about its limits. It can only find watermarks from companies that have signed up to SynthID, so a picture from a model that doesn’t use it, or one that has been heavily edited, will come back as “not detected”. Google warns that this “does not totally rule out AI generation or manipulation”, and that in “very rare cases” the detector can wrongly flag media that was never watermarked.
It also admits the watermark can be beaten: “if someone tries many transformations, they may be able to find one that doesn’t get detected.” Google calls SynthID “one important layer in a broader approach to content provenance”, not a guarantee, and recommends gathering other evidence before deciding whether something is real.
The terms of service ban using the site to probe or break the watermark, scraping it or running it as an automated checker, and getting round its rate limits with extra accounts. Google says uploaded files are processed in real time and deleted once the result comes back.
OpenAI is on board, for images at least
The list of partners matters as much as the tool. OpenAI has adopted SynthID for the media its models generate, so a ChatGPT image can be checked on Google’s site. That is a different approach from its writing: OpenAI has just started hiding its own watermark in ChatGPT’s text in Europe, with a detector only approved researchers can use. Microsoft and Meta still run their own separate systems.
Google has been pushing SynthID well beyond pictures. Last week DeepMind showed it could hide a watermark inside AI-designed proteins that survives in the real molecule.
Why it matters
Fake images, cloned voices and AI video are getting harder to spot by eye, and most people have had no easy way to check. A free, public tool from Google that also covers OpenAI’s and Nvidia’s output is the closest thing yet to a shared checker, but a clean result only means no partner’s watermark was found, not that something is real.
Sources: Google SynthID Detector; TechCrunch.
Latest Google news
