what is SynthID?

SynthID modifies the generation process itself rather than post-processing the output. for text, it nudges the LLM's token-sampling distribution at each step toward a pseudo-random signature pattern keyed to a secret seed. the result is a statistical signal invisible in any single sentence but detectable across a few hundred tokens. for images and video, it perturbs pixel values in ways tuned to survive compression, cropping, and color shifts while staying imperceptible to humans.

Google DeepMind announced SynthID for images in August 2023, expanded it across modalities through 2024, and open-sourced the text watermarker (SynthID-Text) in October 2024 alongside a Nature paper. Google says SynthID is applied by default to Gemini, Imagen, Veo, and Lyria outputs.

it's the most production-deployed watermark today, but only a small fraction of generated content in the wild carries it.

SynthID only marks content from Google's own models. ChatGPT, Claude, Midjourney, Stable Diffusion, Llama, Flux outputs have no SynthID signature. the verifier returning “no watermark” tells you almost nothing on its own. most AI content in the wild today is unwatermarked.

the text watermark is probabilistic, not deterministic. it needs a few hundred tokens to call confidently, and aggressive paraphrasing or translation degrades it. image/video watermarks survive moderate edits (compression, cropping, color shifts) but not all attacks. aggressive downscaling, screenshot-of-screenshot, and AI upscalers all weaken the signal.

Google has not published independent third-party robustness audits. the Nature paper's evaluations are DeepMind's own. treat the published robustness claims as upper bounds, not guarantees.

the common misconception is “if SynthID can't detect a watermark, the content is human-made.” absence of SynthID proves nothing. there's no SynthID-style signal in the file, which is the default for anything from a non-Google model or anything that's been heavily re-encoded.

that gap is why amige. doesn't lean on watermarks at all. a trained model reads the marks a generator leaves and routes each scan to the detectors strongest for it (that routing hint never enters the verdict), then a panel of independent detectors built by different teams weighs in, their reads are fused and calibrated, and it returns “uncertain” rather than guessing when they conflict. see how it works end to end in the machine.