SynthID

What Is SynthID?

SynthID is a watermarking system developed by Google DeepMind, designed to embed imperceptible signals into AI-generated content that survive editing, compression, and reformatting — enabling downstream detection of synthetic media.

SynthID was publicly announced in August 2023 and deployed across Google's Gemini models for text, Imagen for images, Lyria for music, and Veo for video.

How SynthID Works — Text

For AI-generated text, SynthID uses a technique called statistical watermarking. During generation, the model's token sampling distribution is subtly biased using a cryptographic key. This bias is imperceptible to readers but statistically detectable when you have access to the key.

Specifically: at each token position, a pseudo-random "tournament" is run using the key. Tokens that "win" the tournament are slightly favored in sampling. Over hundreds of tokens, this creates a pattern detectable by SynthID's verification system but invisible to humans.

How SynthID Works — Images

For images, SynthID embeds a digital watermark directly in the pixel values using a deep neural network trained to add imperceptible perturbations that survive JPEG compression, cropping, color adjustment, and screenshots. A separate classifier network detects the watermark.

Robustness and Limitations

SynthID text watermarks degrade significantly after heavy paraphrasing or passage through a humanizer tool. Google acknowledges this: SynthID is not robust to adversarial attacks that specifically target statistical watermarks.

SynthID image watermarks are more robust — surviving common image transformations — but can be removed by adversarial perturbation methods.

SynthID also does not work on content generated by non-Google models. It is a provenance signal for Google's own AI outputs, not a universal detector.

Regulatory Context

SynthID is Google's implementation approach for AI watermarking requirements under the NIST AI 100-4 guidance and anticipated EU AI Act obligations. See Standards for the full regulatory picture.