What is SynthID? How Google's invisible AI watermark works
SynthID is Google DeepMind's watermark for AI images, video, audio and text. How it is embedded and detected, which models carry it, and what it cannot tell you.
4 min read
On this page8 sections
Short answer. SynthID is a watermark that Google DeepMind embeds in the pixels of AI images and video, the waveform of AI audio and the word choices of AI text. It is invisible, it is designed to survive cropping, compression and screenshots, and only a detector trained on it can read it. Google adds it to everything Gemini, Imagen, Veo and Lyria produce, and partners have adopted it. It is not metadata, so metadata tools cannot see it or remove it.
Where SynthID came from
Google DeepMind announced SynthID in 2023 for images made with Imagen. Over the following two years it extended the technique to audio, video and text, and Google now applies it across its generative models: Imagen and the Gemini image models (including Nano Banana), Veo video, Lyria music and Gemini's text output. In 2025 Google launched a public SynthID Detector and began opening the technology to other companies, and it has since appeared in output from partners.
The idea is a mark that stays with the content after the file has been stripped of metadata, re-saved or screenshotted, which is where C2PA Content Credentials fail.
How it works
For images and video, SynthID is embedded at generation time by a neural network that adjusts pixel values in a pattern spread across the whole picture. The adjustments are small enough that the image looks identical to one made without the mark. A second network, the detector, has been trained to recognize the pattern and answers with a confidence level: watermarked, possibly watermarked, or not detected.
Because the pattern is distributed rather than placed in one spot, it survives the edits people normally make:
- cropping, as long as enough of the image remains;
- resizing and rotation;
- JPEG compression and re-saving;
- color filters and brightness changes;
- screenshots, which copy the pixels and so copy the pattern.
For audio, the mark is embedded in the spectrogram. For text, Gemini biases its choice of words in a way a detector can score statistically, which is why SynthID text detection works on longer passages and not on a sentence.
What SynthID looks like
Nothing. There is no pattern to see, no logo and no visible noise. The sparkle in the corner of Gemini images is a separate, visible mark; SynthID is in every pixel of the same image whether the sparkle is there or not.
Which content carries it
| Source | SynthID |
|---|---|
| Gemini app images (Nano Banana, Nano Banana Pro) | Yes |
| Imagen through Vertex AI and the API | Yes |
| Veo video, including audio | Yes |
| Lyria music | Yes |
| Gemini text output | Yes, where enabled |
| Partner models that adopted it | Yes; ChatGPT images began carrying it in 2026 |
| Midjourney, Stable Diffusion, Flux, Firefly | No; Firefly uses Content Credentials instead |
Does ChatGPT add SynthID? covers the OpenAI case.
How to check for it
Google's SynthID Detector takes an uploaded image, video or audio file and reports whether a SynthID mark is present, and for images, which parts of the picture are marked. The Gemini app can also answer "was this made with Google AI?" when you upload an image. Nothing else reads SynthID: metadata viewers, C2PA tools and the on-device AI label checker see the file's metadata, not the pixels. Use both to get the whole picture: the checker for the Content Credentials that platforms read, Google's detector for the watermark.
What SynthID does not do
- It does not identify the person who made the image, the prompt, or the date. That is what Content Credentials record.
- It does not mark images from models that do not use it. A clean result from the detector only means "no Google-style watermark found."
- It is not read by social platforms to label posts, as far as any of them have said. Today's automatic labels come from metadata.
- It is not indestructible. Google describes it as robust, not unbreakable, and heavy regeneration of the pixels degrades it at the cost of changing the image.
Can it be removed?
Not by any method that has been shown to work and keep the image intact. Metadata removal does not touch it, cropping and re-saving do not defeat it, and the sites that advertise SynthID removal do not publish detector results. Can you remove SynthID? goes through what is known. HideSlop removes the metadata layer and changes the file's fingerprint; it makes no claim about SynthID.
FAQ
How does SynthID work in one sentence?
A neural network nudges pixel values across the whole image into a pattern that a matching detector can recognize, without changing how the image looks.
Does SynthID survive a screenshot?
Google says yes. A screenshot copies the pixels, and the pattern is in the pixels.
Is SynthID the same as the Gemini sparkle logo?
No. The sparkle is a visible logo in one corner. SynthID is invisible and spread across the image. Both are added to Gemini images on most plans.
Can anyone detect SynthID?
Google's SynthID Detector is public, and the Gemini app can check images. There is no third-party detector, because the detection model is Google's.
Does SynthID affect image quality?
Not visibly. The changes are below what people can perceive, which is the whole point of the design.