Google DeepMind expands SynthID to audio, text, images, and video
Google DeepMind has expanded its SynthID watermarking tool to include AI-generated audio and text, in addition to images and video. SynthID embeds imperceptible digital watermarks directly into AI-generated content produced by Google's generative AI products, including Gemini, Lyria, and Notebook LM. The tool aims to foster transparency by allowing users to check if content was created or altered by Google AI.
Key Takeaways
- SynthID now embeds imperceptible watermarks in AI-generated images, audio, text, and video.
- Google says the watermark survives common edits like cropping, filters, frame-rate changes, lossy compression, MP3 compression, and speed changes.
- Audio watermarking applies to content generated through Lyria and NotebookLM’s podcast generation feature.
- Gemini text watermarking works by adjusting token probability scores during generation.
- Google launched the SynthID Detector earlier this year and is testing it with journalists and media professionals.
Why It Matters
Google is extending provenance tooling across the full stack of its generative products, making it easier to identify AI-generated or altered media inside Gemini, Lyria, and NotebookLM. For streaming and video workflows, the relevant point is that the watermark is designed to survive edits common in production and distribution, including cropping, filters, frame-rate changes, lossy compression, and MP3 compression. Google is also pairing the consumer-facing Gemini check with a SynthID Detector portal for journalists and media professionals. The next concrete signal to watch is whether that detector leaves early testing and becomes broadly available beyond the current waitlist.
Read full article at deepmind.google