Google’s SynthID Detector is now open to everyone
Google opened its SynthID Detector to the public on Tuesday, giving anyone with a browser a free way to check whether an image, video, or audio file was made with AI. TechCrunch reports that the site, which had been restricted to select journalists, media professionals, and researchers since a pilot at Google I/O last year, now accepts uploads from the general public.
The detector takes the formats you’ll actually run into online. Images: JPG, JPEG, PNG, BMP, WEBP, AVIF, TIFF, TIF, and GIF. Video: MP4, MOV, and WEBM. Audio: WAV, MP3, OGG, FLAC, AAC, and M4A. MacRumors notes that you’ll need to sign in with an account to use it, and for now it works globally in English only — a real limitation for a tool whose most urgent users would often be non-English fact-checkers working under deadline.
What’s actually behind the watermark
SynthID is the invisible watermark Google has been embedding in AI-generated content since 2023. It tags output from models like Nano Banana, Veo, and Lyria, plus tools with generation features such as Gemini, Flow, ProducerAI, and Vids. Google says it has now watermarked more than 180 billion images and videos and 240,000 years of audio, and the built-in verification inside Search, the Gemini app, and Chrome already fields more than one million requests a day.
The coalition behind it is bigger than Google alone. OpenAI, Nvidia, and Kakao also support SynthID, and Apple is expected to join soon. OpenAI runs its own verification page for its media, while Microsoft and Meta still rely on their own separate standards — a fragmentation problem that matters more than the launch headlines admit.
Why this matters: read the verdict before you trust it
Here’s what the launch-day briefs gloss over, and it’s the whole story: the site’s own FAQ states plainly that this is not a general AI detector. It can only find watermarks from models that opted into SynthID. A file with “no watermark detected” might simply come from a model outside the coalition — including models from Microsoft or Meta, whose own verification tools have historically struggled to flag even their own companies’ output. Heavy modification can degrade the signal, and the system can occasionally flag content that carries no watermark at all.
It also can’t tell you whether AI created something or merely touched it up, and a clean result is not proof an image is real. For journalists and anyone verifying viral media, the playbook is straightforward: upload the highest-quality original file you can find, treat every verdict as one data point among several, and never mistake “no watermark detected” for “definitely authentic.” Provenance is a process, not a button.
Meanwhile in Apps & Tools:
For a different take on trustworthy AI, check out the underdog AI assistant that runs entirely on-device with no subscription — proof that the most reliable tools are sometimes the ones that never touch the cloud, and that the best software is the kind you can actually inspect.
