Anthropic reveals Claude text watermarking details, built into latest models

新智元 · wechat · 2026-08-15

Anthropic has publicly detailed its text watermarking technique for Claude, now built into all latest models. The watermark subtly alters word selection using a key-derived random sequence, with no impact on output quality per Google's A/B test. It's driven by the EU AI Act. The watermark works in translation and code comments but is ineffective for proofreading, short texts, and factual content. Research shows paraphrasing tools can remove 98.3% of watermarks at a cost of $0.04. A detection API is coming but details are pending.

Related event: Anthropic Implements Universal Text Watermarking for Claude(8 posts)→

Original post →

More from Safety

Safety channel →