What Google's SynthID Detector post opens

What happened

On October 7, 2026, Google DeepMind published a post by Pushmeet Kohli opening a SynthID Detector to the public. The post says the detector is available globally, in English, for images, video, and audio. It names OpenAI, NVIDIA, and Kakao as partners, and says Apple is coming soon. It says SynthID watermarks have been applied to more than 180 billion images and videos and to 240,000 years of audio. It also says built-in checks in Search, Gemini, and Chrome already handle over 1 million requests a day. Those sentences are the company's account. The front page that files them is Futureweb.

What the post actually ships

The post ships a way to ask whether a file carries a SynthID watermark. It does not ship a verdict that an unmarked file was made by a person. A reader who treats a negative result as proof of human authorship is adding a sentence the post does not contain. Ars Technica's account of the rollout adds a login and a daily quota, and says the public site reports whether SynthID was used. That login and that quota are the outlet's description of the tool, not a line this page is taking from the short company post. The desk filing the post is Artificial intelligence.

What the outlets add

Ars Technica, 9to5Google, and CNET all reported the same October 7 opening: a public detector, English, worldwide, covering images, video, and audio. They do not replace the partner list. OpenAI, NVIDIA, Kakao, and the soon-Apple line stay in Google's post. Where an outlet adds a file-type list or a quota, that detail stays with the outlet. This page does not merge those extras into the company sentence. A standing note on what a watermark can and cannot show a reader is a watermark a reader can test.

Why it matters on this desk

This desk files what a reader can check about a model artifact. Here the check is a watermark detector with a named scope: images, video, and audio, in English, globally, as of the October 7 post. The partner names are a second check. They say whose files the detector is meant to recognize, and they are a list, not a claim that every generated file on the internet carries the mark. What on-device inference keeps, which is a different boundary, is what on-device inference keeps.

What a missing mark does not prove

A detector that looks for one vendor's watermark can say the mark is present or that it did not find the mark. It cannot, from a miss alone, say a camera made the file, a person painted it, or another lab's model made it. Files can be cropped, recompressed, or produced by a system that never wrote the mark. The post's own volume figures, 180 billion images and videos and 240,000 years of audio, describe how much Google says it has marked. They do not describe the unmarked rest of the web. A question about who holds a key, which this detector is not, sits on Privacy.

What the daily request figure is

The figure of over 1 million requests a day is Google's description of checks already running inside Search, Gemini, and Chrome. It is not a census of how often the new public detector page is used, and it is not a promise about accuracy. Ars Technica's separate note of a login and a quota applies to the public website the outlet tested. A reader can hold both: the company post describes built-in checks at that volume, and an outlet describes limits on the new page. Neither number is a measurement this paper ran.

What to watch next

The page changes when Apple is actually in the partner list, when the public tool's quota is written on Google's own page, or when the detector's result language changes from a watermark check to something else. A later post that adds a language besides English changes the scope sentence. A later post that only repeats the 180 billion figure does not. The October 7 post is the record for the opening.

What would change the record

A Google page that states the login rule, the daily cap, and the file types would let this brief stop attributing those limits to Ars Technica. A retraction of the partner names would change the list. An independent test that the public tool mislabels unmarked files would be a different story, and it is not what the October 7 outlets filed. Until then the checkable facts are the date, the three media types, the partner names, the two volume figures, and the built-in request figure.

What stays a watermark check

What stays specific is a public detector for one watermark family, opened in English on October 7, 2026, with partners named in the company post. It is not a general test of whether a file is synthetic. It is not a claim that every image on the web was marked. The outlets agree on the opening. They do not agree, in the text this page is using, on every limit of the website. Those limits stay attributed. The company sentences do not. 9to5Google's report is the third account of the same October 7 opening, used here so the detector is not a single-outlet headline.