Anthropic is introducing invisible text watermarking for future Claude models to comply with the EU AI Act. Built on Google DeepMind’s SynthID-Text technique, the technology alters the subtle statistical randomness used when selecting words.
This creates a verifiable pattern detectable only via a verification key, without changing readability, latency, output quality, or token costs. The watermark contains no personal user data and is applied globally at launch. While highly effective on long prose, its impact is minimal on exact outputs like code or light proofreading. A detection API will follow soon.
Source: Anthropic news