Google opens SynthID Detector to everyone, for partner watermarks as well as its own

What did Google open on 7 October?

Pushmeet Kohli, Google DeepMind’s vice president of science and strategic initiatives, posted that SynthID Detector is now available to everyone, globally in English. Last year the same product was a waitlist for journalists, media professionals and researchers. The 20 May 2025 announcement said testers would get access first, “before making it available more broadly.”

Kohli’s October figures are vendor counts: more than 180 billion watermarked images and videos, and 240,000 years of audio, since SynthID launched in 2023. Built-in checks in Search, the Gemini app and Chrome “now regularly handle over 1 million requests daily.” Those are Google’s numbers. We did not audit them.

The DeepMind SynthID product page was updated to match: after testing with journalists, “we’re expanding access to everyone,” with a try-now link. We opened synthid.com on 7 October. The public page asks for a file upload and states that it detects media from companies that have adopted SynthID. A public portal is a product change. A watermarking stack that already existed is not. For how to keep those apart, see our note on reading AI company announcements.

How does the detector work, and what files does it accept?

SynthID embeds an imperceptible mark in the pixels of images and video and in the frequency components of audio. Unlike C2PA Content Credentials, the mark is not a sidecar metadata block that a conversion can strip. Google designed it to survive cropping, filters, frame-rate changes, MP3 compression and speed changes. The Detector looks for that intentional signal. It does not try to guess “AI-looking” artifacts the way a generic classifier does.

The site’s FAQ, inspected 7 October, lists three current content types: images (presence of a mark), video (which segments carry it) and audio (which segments carry it). It does not list text on the public Detector, even though SynthID-text already exists in Gemini. The May 2025 post had promised video and text detection “in the coming weeks” for the early portal. Treat today’s public tool as image, video and audio unless Google updates that FAQ.

TechCrunch, reporting the same day, listed common image, video and audio containers the site accepts. That list is TechCrunch’s, not a specification we downloaded from Google. We did not upload test files.

Who can use it, and what does a result mean?

Kohli says anyone can use it, in English, worldwide. The Verge reported that use requires logging into synthid.com. The page we opened asked for an upload; we did not create an account or run a file through it.

Two outcomes are defined. “Watermark detected” means a SynthID mark is present and the file was likely generated or edited by a SynthID-enabled model. “Watermark not detected” means no confident signal. The FAQ says that does not totally rule out AI generation or manipulation. Misses can come from an unmarked model, pre-adoption content, or heavy edits. False positives are possible “in very rare cases.” Google asks for the highest-quality file available.

Partners named today are OpenAI, NVIDIA, Kakao, and Apple “soon.” The 2025 post had already named NVIDIA Cosmos preview video and GetReal Security. OpenAI’s later provenance work, documented in its Content Provenance API guide inspected on 7 October, checks images for C2PA and SynthID and audio for SynthID through POST /v1/content_provenance_checks. A not_detected result there also does not prove the file is human. That sits next to OpenAI’s EU text-watermark plan, which is a different mark, for text, and is not what synthid.com checks today.

How does this sit next to Gemini, Chrome and OpenAI’s own checker?

Google already ships in-product checks. A user can upload a file in Gemini, or ask Search or Chrome whether something was made with Google AI. The Gemini help page, checked 7 October, still says Gemini can currently only recognize content created by Google AI tools, “while other companies have started to adopt SynthID watermarks.” That is a narrower claim than today’s Detector. Readers should not assume the Gemini chat check and synthid.com have the same partner coverage until Google updates that help page.

TechCrunch says OpenAI maintains its own verification site. OpenAI’s developer docs, which we did open, describe image and audio checks plus an API; text verification is “currently available only to approved organizations.” We received HTTP 403 from openai.com/research/verify/ on 7 October and are not quoting that page as opened HTML.

TechCrunch lists Nano Banana, Veo, Lyria, Gemini, Flow, ProducerAI and Vids as Google products that watermark outputs. Google’s Nano Banana 2.1 model card already documents SynthID on generated images. Last month’s SynthID Bio protein watermarks are a different method, for sequences and structures, and are not what this portal reads.

What should readers not assume from a watermark result?

A detected mark is a provenance signal, not a statement that the picture is false, the audio is harmful, or the publisher is the model vendor. A miss is not a clean bill of health. TechCrunch notes that Microsoft and Meta use their own watermarking schemes; those files will typically miss on SynthID Detector.

There is still no single industry watermark. The Detector FAQ says standards take time and that watermarking should sit next to metadata standards. C2PA can travel with a file and then vanish on export. SynthID can survive some edits and still be stripped by enough transformations. Neither identifies a user.

We have not tested removal attacks on this portal. DeepMind’s SynthID Bio paper documented removal paths for protein marks. Media watermarks are a different method, but a published detector is also a target. Google’s FAQ already says enough transformations may defeat detection.

What remains limited after today’s launch?

The public site is English only. The FAQ we inspected covers image, video and audio, not text. Apple is named as forthcoming, not shipping. The partner list is Google’s; we did not verify mark presence model by model. Google did not publish a false-positive rate. The 180 billion and 1 million figures are Google’s counts. Treat 7 October as a door opening on a mark that was already being written, not as a new way to prove a file is real.