Putting an invisible watermark in AI-generated text — without distorting a single word choice. The clever mechanism is now public. Title: How Claude's text watermark works URL: Claude is adding invisible text watermarking to future models to comply with the EU AI Act. Three points worth a closer look. 🎲 Highlight 1: It changes the "source of randomness," not the words Watermarking sounds like swapping words, but this method doesn't. It only replaces the source of randomness used during generation, deriving it from an encoding key and preceding words. Like using digits of pi instead of dice in Monopoly: readers can't perceive it, yet it can be verified after the fact. The technique is SynthID-Text, published by Google DeepMind in Nature. ⚡ Highlight 2: Near-zero impact on quality, speed, and cost Internal tests and DeepMind's controlled studies found no statistically significant difference between watermarked and unwatermarked text. Speed impact is negligible, no extra tokens are needed, and pricing is unchanged. The watermark carries no info traceable to users, orgs, or specific chats. 🔍 Highlight 3: Honest about what it can't do Longer passages detect more reliably, but short samples are hard. It can't watermark fact-heavy spans or exact outputs like equations and functional code where there's no equivalent alternative. Light edits may survive, but a full rewrite erases it. A detection API is coming soon. A pragmatic take on transparency that balances regulation with production quality. #AIWatermark# #EUAIAct#