Google Opens SynthID Detector Globally With Partner AI Content Checks

SynthID Detector recognises only its partners' watermarks, so a clean result proves nothing about who made a file

Google has opened its SynthID Detector to anyone, worldwide, in English, according to Unite.AI, Search Engine Journal and 9to5Google. Google DeepMind's Pushmeet Kohli announced it in October 2026. Upload an image, video or audio file to synthid.com and it checks for watermarks from Google, OpenAI, NVIDIA and Kakao; Apple is promised, undated. Google says it has watermarked over 180 billion images and videos and 240,000 years of audio since 2023, and that built-in checks in Search, Gemini and Chrome handle over 1 million requests daily.

Can SynthID Detector tell me whether a file is AI-generated?

Only partly. It recognises watermarks from Google and named partners, so a positive result is meaningful, but a negative result cannot separate human-made files from output by tools outside the scheme.

Google's own FAQ says: "This is not a general AI detector." A hit is decent evidence that a participating tool touched the file. A miss proves nothing, since any other generator's output comes back blank, so 9to5Google's framing of the tool as a human-made check is a misreading. Search Engine Journal adds that a hit only says the file was made or edited with Google AI, so a retouched photo and a synthetic one share a label.

Does it help with AI-written text?

Not through the portal as described. It checks images, video and audio; the coverage describes no text check there, though DeepMind documents a text watermark for Gemini app output.

DeepMind's SynthID page says the text watermark works by adjusting the probability scores a model gives candidate words. The coverage doesn't say whether a text mark survives paraphrasing, or when a public text check might arrive. Nothing here changes how you screen written work today.

How far should I trust Google's scale figures?

Treat them as output counts, not accuracy. Google supplies the totals itself, and the coverage cites no independent measurement of how often the detector hits or misses.

In May 2026 Google reported over 100 billion watermarked images and videos, 60,000 years of audio and 50 million Gemini-app checks. The latest totals are 180 billion and 240,000 years. Audio grew far faster than images and video, and nobody says over what window, or whether partner output is counted. Every figure is output or usage volume. None speaks to hit rate or false alarms.

What should I ask before relying on a SynthID result?

Ask which generators are covered, what the result wording means, what detection rates exist for edited files, and whether a text check is coming. Until answered, treat a result as one data point.

Start with Apple, where the coverage disagrees. Unite.AI and Search Engine Journal list it as still to come. 9to5Google lists it among partners already, while also saying Apple Intelligence images will carry the watermark soon. 9to5Google also implies everything those companies generate is marked, which is broader than Search Engine Journal's claim about consumer products.

Our earlier piece on OpenAI's findings about Russian and Iranian fake news operations suggests why the gap matters. Our inference is that a check seeing only cooperating vendors' media does little against operators who pick other tools. We also covered Google hiring to generate AI content for Discover; ask whether that output is marked.

Frequently asked questions

Who can use SynthID Detector?

Anyone, globally, in English. Google's post references no other languages.

Which companies' output does it recognise?

Google, OpenAI, NVIDIA and Kakao today. Apple is expected, with no release date.

What should I upload?

The best-quality copy you have. The site says that gives the most accurate result.

Sources: Unite.AI, Search Engine Journal, 9to5Google.