Krux
March 29, 2026
Google Watermarks Every Second of AI Voice Output
Published: March 29, 2026 at 2:51 AM
Updated: March 29, 2026 at 2:51 AM
100-word summary
Google just launched Gemini 3.1 Flash Live, its fastest voice model yet, with one unusual feature: every audio response gets watermarked. The move aims to curb deepfakes as voice AI gets convincing enough to handle customer service calls and process returns without sounding robotic. The model scored 90.8% on ComplexFuncBench Audio and ships with lower latency than previous versions. It's live now in over 200 countries through Google Search and Gemini Live, with API access in preview for developers. What you can do today that you couldn't yesterday: build a voice agent that actually understands when a customer is annoyed, not just what they're saying.
What happened
Google just launched Gemini 3.1 Flash Live, its fastest voice model yet, with one unusual feature: every audio response gets watermarked. The move aims to curb deepfakes as voice AI gets convincing enough to handle customer service calls and process returns without sounding robotic. The model scored 90.8% on ComplexFuncBench Audio and ships with lower latency than previous versions. It's live now in over 200 countries through Google Search and Gemini Live, with API access in preview for developers.
Why it matters
What you can do today that you couldn't yesterday: build a voice agent that actually understands when a customer is annoyed, not just what they're saying.