the watermarking is nothing else for the end user than fixing a random seed and making this sampling kind of deterministic
这句拆掉了「水印会让输出变差」的直觉恐慌：水印只作用在采样层，不改模型权重，也不必重训——它是在若干等概率的候选词里按密钥挑一个。代价不在质量而在多样性和可复现性。但也要看到反面：正因为它只是采样偏置，抗改写能力天然有限。
4 Matching Annotations
- Aug 2026
-
magazine.sebastianraschka.com magazine.sebastianraschka.com
-
- Jul 2026
-
blog.google blog.google
-
Gemini Omni and Nano Banana 2 Lite use [SynthID](https://deepmind.google/blog/identifying-ai-generated-images-with-synthid/) watermarking. You can verify AI content through the Gemini app, Gemini in Chrome or Search.
文章提到使用SynthID水印来验证AI内容，需要核查这一水印系统的有效性和可靠性。
-
- Apr 2026
-
blog.google blog.google
-
All audio generated by Gemini 3.1 Flash TTS is watermarked with SynthID. This imperceptible watermark is interwoven directly into the audio output, allowing the reliable detection of AI-generated content to help prevent misinformation.
令人惊讶的是：该模型使用名为SynthID的不可察觉水印技术，将水印直接编织到音频输出中，以便可靠地检测AI生成的内容。这种技术对于防止AI语音被用于传播虚假信息至关重要，但大多数用户可能并不了解这种隐形水印的存在和工作原理。
-
- Jan 2021
-
theintercept.com theintercept.com
-
Journalists should also be wary of publishing raw audio leaked from Zoom meetings, particularly if the source is not sure whether audio watermarking was enabled or not.
-