Anthropic · Aug 14
Anthropic built an invisible watermark that flags text written by Claude
Anthropic detailed an invisible watermark that subtly biases Claude's word choices so AI-written text stays detectable after copy and paste, with no hidden characters and no hit to cost, speed, or quality. The company tied it to the EU AI Act content-marking rule that took effect August 2, 2026, while admitting detection weakens on short, factual, or code output.
▸ The MSP Angle
How can I tell if content a client sent us was written by AI?
Content-labeling rules are now live in the EU, and clients will ask whether the marketing, code, and documents moving through their business are machine-written. Vendor watermarks only cover their own model's output and break on short or edited text, so the durable answer is a governance policy: log which AI tools staff use, where the output goes, and which work needs disclosure. That policy is a billable service, not a checkbox.
SiliconANGLE · Aug 13