On October 7, 2026, Google opened a door it had kept half shut for more than a year. The SynthID Detector, a website that searches an image, a video or a sound file for an invisible signature, became available to anyone, anywhere, in English. Until then it had belonged to a small circle of journalists, media professionals and researchers, who were given it for testing at Google I/O the year before, as Ivan Mehta reported for TechCrunch. Now the portal at synthid.com is open to the public, with conditions.

The signature itself is older. Google DeepMind introduced SynthID in 2023, first for images and later for text, video and audio. In a picture the mark lives in the pixels; in a recording it is woven into the waveform, according to accounts of the launch. Pushmeet Kohli of Google DeepMind explained the purpose in a company blog post: “As AI tools become better and easier to use, having clear context about the content you encounter online helps you make informed decisions about what to trust.”

Those who sign, and those who are seen

Every image, clip and song that leaves Google’s own workshops carries the mark: the Nano Banana, Veo and Lyria models, and generating tools such as Gemini, Flow, ProducerAI and Vids. Google, which has lately been handing out AI video for free, signs its own output. Other companies have joined. The detector also recognises media made with tools from OpenAI, Nvidia and Kakao, Juli Clover wrote for MacRumors, and Apple plans to add support later this year. OpenAI announced in May that it would put the SynthID watermark in its images alongside C2PA metadata, and it runs its own verification page for content from ChatGPT, its API and Codex.

The site takes almost any format: JPG, JPEG, PNG, BMP, WEBP, AVIF, HEIC, HEIF, TIFF, TIF and GIF for images; MP4, MOV and WEBM for video; WAV, MP3, OGG, FLAC, AAC and M4A for audio. The same checks are already built into Search, the Gemini app and Chrome, and Google says people send them about a million verification requests a day.

The amount already marked is very large, though the reported figures differ. One report puts it at more than 180 billion images and videos watermarked since 2023, plus 240,000 years of audio. Others give over 100 billion images and videos and 60,000 years of audio for Gemini alone. The two sets of numbers probably measure different things.

The ration

Anyone may use the detector, but not without limit. A visitor must first sign in with a Google, OpenAI or Apple account, Ryan Whitwam noted at Ars Technica, and Google sets the allowance in a careful phrase:

a daily usage quota of approximately 10 image, video, and audio checks.

Approximately ten. Google engineers have told Ars the limit is needed so that people cannot use the checks to build tools that strip SynthID out. The vagueness probably has a reason too: the system may lock a user out sooner if they submit many very similar images, which could be a sign that someone is testing a way around the mark.

A reflection, not a finding: every authority that guards a secret ends up rationing access to it, and it treats the citizen with a question and the forger with a workshop as the same person. Here the citizen gets ten questions a day so the forger cannot get a thousand.

What the lamp cannot see

Google itself says the website is not a general AI detector. It finds only media carrying SynthID, and the site will carry a disclaimer to that effect. Large areas stay dark. Meta uses different watermarking technology. Some model makers use none at all. Open-weight models, which people run on their own hardware, produce increasingly realistic images that do not even carry simple metadata labels. Microsoft and Meta have their own standards for marking and verifying content, but those tools often fail to recognise what their own makers’ models produce. Google’s detector also cannot tell AI-created media from AI-edited media, and watermarks can be removed. Ars considers it plausible that someone will find a way to strip SynthID despite Google’s efforts, which would make existing fakes look more genuine.

The road between the generator and the viewer also loses information. Researchers who tested images shared on X found that none of the genuine photographs were wrongly flagged as carrying SynthID, but only 36 of 200 images from OpenAI’s GPT-Image-2, or 18 percent, were flagged. Every C2PA check came back negative, because X strips metadata when files are uploaded.

Whitwam argues the better defence may run the other way: label real images, video and audio with cryptographic technology such as C2PA at the moment they are captured. Google’s Pixel phones are among the few devices that already do this. Until that becomes common, his verdict is plain: “SynthID is the best bet you’ve got to identify AI fakes.”

Google says it is “continuously working on extending our partnerships to allow detecting more generated media.” For now, the most a person can learn on any given day is whether one of a small group of companies signed the thing in front of them, approximately ten times.