Generated content watermarking

Nature, October 23, 2024: Google DeepMind published SynthID-Text, a watermark that biases word sampling, and open-sourced the code. It had been tested live on nearly 20 million Gemini responses, with no quality difference users could detect. Two families of technique answer the same question without being the same thing. A watermark embeds an invisible statistical signal in the content itself, which a detector can recover. Provenance attaches cryptographically signed metadata to the file, describing how it was made and later modified. On paper the second approach is stronger, since it rests on a signature. In practice it is fragile: a screenshot or a recompression wipes the metadata. A watermark survives that kind of handling but gives way to a deliberate attack. Neither solves the problem alone, and combining them is the only serious path. The European regulation requires providers of systems that generate synthetic content to apply machine-readable marking, and requires deployers to disclose deep fakes. The limit is built in: these mechanisms prove that content was generated, never that unmarked content is authentic.