Definition:
SynthID is Google DeepMind’s invisible watermarking tool that embeds a hidden, imperceptible digital signature directly into AI-generated content (images, audio, text, and video) at the moment of creation — so you can later verify whether something was made by Google’s AI.
Think of it like invisible ink on a banknote: the watermark is there, it doesn’t change how the bill looks, but under the right scanner, you can prove it’s authentic (or in this case, AI-generated).
How It Works (One Sentence per Modality) Content Type Watermarking Method Images & Video Tiny imperceptible changes to pixels in every frame — survives cropping, filters, and compression. Audio Inaudible signal embedded in the waveform — survives MP3 compression and speed changes. Text Subtly tweaks the probability of which word the model picks next during generation — invisible pattern to humans, detectable by the SynthID detector.
Example:
Imagine you ask Gemini: "Write a short poem about the ocean."
- During generation, SynthID subtly nudges the model's word choices. For the line "The waves crash upon the..." the model might normally have a 60% chance of picking "shore" and 30% for "sand." SynthID might flip those odds slightly — say, 55% for "sand" and 35% for "shore." To a human, the poem still sounds perfectly natural. But the pattern of those tiny nudges forms a hidden signature.
- Later, someone uploads that poem to the SynthID Detector. The detector analyzes the statistical pattern of word choices, compares it against the expected watermark pattern, and returns: "Yes, this text was watermarked by SynthID — likely generated by Google AI."
- The key insight: If the poem is short and factual (like "What's the capital of France?"), SynthID has very little room to nudge probabilities without changing the answer — so watermarking is weaker. But for creative, longer outputs (essays, scripts, varied responses), the watermark is robust.