SynthID: How invisible watermarks mark AI-generated images, video and audio
SynthID hides an invisible signal inside AI-made images, video, audio and text so a detector can later recognize it. Here is how it works, who uses it and where it falls short.

SynthID is a watermarking system from Google DeepMind. It hides a signal inside AI-generated images, video, audio and text while the content is being created. People cannot see or hear the signal, but a matching detector can find it later and say the content came from a participating AI tool.
It is the mark Google says is on every output of Genie 4, the world model it just opened to developers. Our report on Genie 4 in the Gemini API notes the watermark but not how it works.
A signature in the ink, not a stamp on the page
A visible logo in the corner of a picture can be cropped out in seconds. An invisible watermark works differently. Think of a banknote: the paper carries a security feature that is part of the material itself and shows up only under the right light.
SynthID follows that idea. According to Google DeepMind, it embeds the watermark directly into AI-generated images, audio, text or video, and the mark is imperceptible to people. Because it lives in the pixels, sound or word choices, there is no label to delete.
How the mark is added for each kind of media

The signal has to be built differently for each format, and Google has published details for some formats more than others.
For text, DeepMind explains that a language model writes one token (a word or word fragment) at a time and scores many candidates. SynthID adjusts those probability scores in a patterned way. The text reads normally, but a detector that knows the pattern can test whether the word choices lean toward it.
For images, a 2025 Google paper calls SynthID-Image a deep learning system that invisibly watermarks AI imagery. The sources we read do not describe the embedding method in detail. Google says video is covered as well, and the paper counts watermarked video frames alongside images.
For audio, Google says the watermark is inaudible and is applied to audio from Lyria and the podcast feature of NotebookLM. In every case, the mark is added at the moment of creation rather than afterward.
Where it came from and how far it has spread
The text version was the first to get a peer-reviewed description. Sumanth Dathathri and colleagues at Google DeepMind published “Scalable watermarking for identifying large language model outputs” in Nature in October 2024. It describes a sampling scheme called tournament sampling.
On May 20, 2025, Pushmeet Kohli of Google DeepMind announced the SynthID Detector, a portal where people upload a file to check for the watermark. The post said over 10 billion pieces of content had been watermarked across Gemini, Imagen, Lyria and Veo, and that the text watermarking had been open-sourced.
In October 2025, a 26-author team led by Sven Gowal described SynthID-Image. It reports more than ten billion images and video frames watermarked across Google services. It also describes SynthID-O, a variant offered through partnerships.
DeepMind’s product page now says the Detector is being expanded to everyone. It says users can check whether a file was made with AI from Google or partners, naming OpenAI, NVIDIA and Kakao, with Apple listed as coming. Gemini, Search and Chrome can also run the check on request, according to the same page.
What a watermark can and cannot prove
Google says SynthID is designed to hold up under cropping, filters, frame-rate changes and lossy compression for images and video. For audio it lists added noise, MP3 compression and changed playback speed. The company has not published accuracy figures or false-positive rates in the pages we read, so these are design claims rather than measured guarantees.
Several limits follow from how the system is built:
- It only proves a positive for known generators. A missing watermark does not mean a file is real. It may come from a model that never added one.
- Heavy edits can erase it. Text is the weakest case. Secondary reports of research from 2025 say paraphrasing, back-translation and similar rewriting can degrade the text watermark sharply. We did not read that study directly.
- Detection depends on the vendor. Checking for a given watermark generally needs the detector of the company that made it.
Kohli has described SynthID as a building block rather than a complete answer, according to press coverage of the Nature paper.
Watermarks, signed labels and what to watch
SynthID is often compared with C2PA Content Credentials, a standard that attaches a cryptographically signed record of how a file was made. The signed record carries more detail, but a screenshot or re-encode can strip it. A watermark stored in the content itself is harder to strip but carries only a yes or no signal. Many observers treat the two as complementary.
For developers using Genie 4, the practical point is that generated worlds carry the mark by default, as Google describes it. That helps anyone who later needs to show a clip came from a Google model, though it says nothing about clips from other tools.
Three things are worth watching. One is whether other labs adopt SynthID or a compatible scheme, since Google’s list of partners is still growing. Another is whether Google publishes measured detection rates, including for edited or re-generated content. The last is whether the Detector’s wider rollout gives ordinary users a reliable way to check files, or mainly a signal that works inside Google’s own ecosystem.
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Daniel ReedAI Research, Safety & Ethics AnalystDaniel Reed currently works as an AI Research, Safety & Ethics Analyst at Aiholics, writing about how changes in artificial intelligence are affecting and will affect scholarship,…



