SynthID
Updated Jun 1, 2026
SynthID is Google DeepMind's watermarking system for AI-generated content. Unlike C2PA, which is metadata attached to the file, SynthID is embedded in the pixel values themselves — imperceptible to a human, statistically detectable by a partner. Strip the file metadata, screenshot the image, even re-compress it through several rounds of social media uploads, and the watermark survives.
The project launched in beta in August 2023 for images generated by Imagen on Vertex AI, expanded to audio (Lyria, the music model behind YouTube's Dream Track) in 2024, and reached text and video in 2025. By 2026, SynthID is the default watermark on every output of Google's Imagen 3, Veo 3, Lyria 2, and Gemini image generation.
How it works at a high level
For images, the watermark is encoded as imperceptible adjustments to pixel values across the frame. For audio, it is embedded in the spectrogram. For text, it is statistical bias in token selection. In all three cases, the encoder and the decoder are paired models — the decoder is trained to recover the watermark signal at high confidence, even after the content has been edited, cropped, filtered, or transcoded.
What it survives
- Screenshotting at any resolution above ~256 pixels per side.
- JPEG compression down to roughly quality 60.
- Standard Instagram, TikTok, and X re-encoding.
- Audio re-recording through a microphone (with degraded confidence).
- Basic crops and color edits.
What breaks it
- Heavy upscaling or downscaling through a different generative model.
- Adversarial removal pipelines (specifically trained to strip the watermark).
- Extreme crops that remove most of the frame.
Who can read it
SynthID is detectable by Google's own infrastructure (YouTube's AI-content checker, Google Search, the SynthID Detector tool) and by partners with access to DeepMind's API. Meta is widely understood to use SynthID detection as a fallback signal when C2PA metadata has been stripped, although Meta does not publish its full detector stack.
For creators, the practical implication: pasting a Gemini-generated thumbnail into a Reel will almost certainly trigger the AI Creator label on Instagram, even if you screenshot the image first. Plan for it rather than trying to strip it.
Related terms
Compliance
C2PA Metadata
Cryptographically signed metadata embedded in media that records origin and edit history. Adopted by OpenAI, Adobe, Microsoft, and (in 2026) Midjourney. The signal Meta uses to auto-trigger the AI Creator label.
Compliance
AI Creator Label
Instagram's opt-in label (launched May 4, 2026) that marks an account as primarily AI-generated content. Auto-detection covers C2PA metadata, watermarks, and audio fingerprinting.
Read more