Google made its SynthID Detector free to anyone in the world on October 7, 2026. People can now upload an image, video, or audio clip to the site and check for a hidden watermark that shows the file came from an AI tool. The checker reads marks from Google and partners OpenAI, Nvidia, and Kakao, with Apple expected to join.
The launch was announced by Pushmeet Kohli, vice president of science and strategic initiatives at Google DeepMind, on the company's blog. Until now, the detector was open only to journalists, media professionals, and researchers, a group that had been testing it since May 2025. The watermark sits inside the file itself, in the pixels of a picture, the frames of a video, or the shape of a sound wave. Because it is not stored in a file's hidden data fields, it can survive cropping, filters, and compression. Checking a file at synthid.com requires signing in with a Google, OpenAI, or Apple account.
Google says it has watermarked more than 180 billion images and videos since 2023, along with the equivalent of 240,000 years of audio. The same technology already runs inside Search, the Gemini app, and Chrome, which together handle over 1 million checks a day. On the Google side, the mark is added by the Nano Banana, Veo, and Lyria models and by the Gemini, Flow, ProducerAI, and Vids tools. Google also backs C2PA Content Credentials, a separate standard that records how a file was made or edited.
The tool has one firm limit, stated plainly in its own help page: "This is not a general AI detector." It finds only the marks of companies that chose to adopt SynthID. Many AI systems, including open models anyone can download, add no mark at all. So a clean result proves nothing on its own. The checker also cannot tell a fully AI-made file apart from a real photo that was later edited with AI.
Beyond the launch itself, the harder question is whether watermarks can carry the weight being placed on them. Sam Gregory, executive director of the human rights group WITNESS and a longtime voice on synthetic media, has warned that a patchwork of competing standards risks confusing the journalists, activists, and ordinary users who must decide what to trust. He argues that watermarking alone cannot solve the problem and has to sit alongside provenance records and policy. Researchers point the same way: a 2026 study found that simple edits and paraphrasing could strip watermarks from text, and other work has shown image marks can be removed with targeted attacks.
That gap is the thing to watch. A checker that reads only its makers' marks is useful mainly when the people spreading fakes want to be honest, which the people spreading harmful fakes usually do not. Its real value grows only if far more companies add the mark and a quick check becomes routine. For now, Google's site answers a narrow question well, but it cannot tell you whether what you are looking at is true.
