Anthropic announces text watermarking for future Claude models
Anthropic announced that future Claude models will add an invisible text watermark to meet EU AI Act content-marking requirements, without adding tokens, cost, or user-identifying information.
What happened
Event details
Anthropic described a watermark based on SynthID-Text that changes low-stakes word-selection randomness without adding hidden characters. Detection is limited for short, factual, and code-heavy passages. Future Claude models will use the watermark globally at launch, older models will be covered during a transition period, and a detection API is planned. Supported generated files use C2PA metadata instead.
Assessment
Why it matters
A compliance-oriented product adjustment scored 38 under impact-v1: technical 1, market 0, ecosystem 1, policy/social 2, scope 2, duration 2.
Availability
Access notes
Future Claude models are planned to use the watermark globally at launch. Older models will be covered over the coming months; the detection API is not yet available.