Google says it has watermarked 240,000 years of AI audio.

Michael Smith ran AI-music streaming fraud undetected for 7 years and got 18 months this week.

SynthID reads imperceptible watermarks that models embed at generation, so it detects only Google, OpenAI, Nvidia, Kakao and Apple output.

OpenAI and Anthropic warn their text watermarks work poorly once a human edits the output.

The European Union’s AI Act already requires generative products to make outputs identifiable by automated systems.

How each outlet framed it

drawn from 150+ reports worldwide

The Independent
contextualizes detector rollout with 54-year-old's 7-year AI music streaming fraud sentencing (18 months, millions in earnings)

Sources: The Independent