What Is SynthID? Google’s Invisible AI Watermark Explained

SynthID is Google DeepMind’s imperceptible watermark for AI-generated content. Unlike a corner logo or a text label, it is embedded directly into the pixels of an image at generation time, invisible to the human eye but readable by a compatible detector. It exists so platforms and people can tell whether content was produced or edited by AI, and it is deliberately hard to strip out. This explainer covers what it is, how it works, where it shows up in 2026, and how detection actually happens.

#How SynthID is embedded

Google DeepMind describes SynthID as a watermark added directly into content as it is generated. For images, the signal is encoded into pixel and frequency-domain patterns rather than stamped onto a visible region. Two properties matter most: it is imperceptible, so the image looks unchanged to a viewer, and it is redundant, meaning the watermark information is repeated across the whole frame rather than sitting in one place.

That redundancy is the reason SynthID is robust. Because the signal is everywhere, cropping a corner, resizing, re-compressing, or converting formats removes only some copies while leaving many intact. This is exactly why it behaves so differently from a logo you can crop away, a point we test directly in whether cropping removes SynthID.

#How robust is it, really

These figures are approximate and depend on the image and the detector, but the shape is consistent: routine edits barely dent SynthID, while rewriting the pixels does the real damage. Independent research such as the UnMarker project has reported regeneration-based attacks succeeding around 79% of the time; Google disputes those results. No watermark is truly permanent, but SynthID is engineered to be far stickier than anything visible.

#Where SynthID shows up in 2026

• Images from Google’s own generative tools, including the Gemini app and Imagen.
• Since 2026, images from OpenAI tools such as ChatGPT and DALL·E, which now embed SynthID as well.
• Increasingly, AI video and audio, where SynthID has been extended beyond still images.

The practical implication is that a clean-looking AI image from a major tool likely carries a SynthID signal even though you cannot see it. If your image came from ChatGPT or DALL·E specifically, our note on removing a ChatGPT or DALL·E watermark is the relevant read.

#How SynthID is detected

Detection is effectively Google-only. Google runs a SynthID Detector portal, and the Gemini app can flag content it recognizes, but there is no open public API for third parties. That is an important limitation: the “AI detectors” you find elsewhere return likelihood scores based on their own models, not a cryptographic verification of a SynthID watermark. For a hands-on walkthrough, see how to check an image for SynthID, and verify provenance separately at contentcredentials.org/verify.