Claude AI Watermarks Text for Authenticity in Cutler Bay

Cutler Bay residents may soon have a clearer way to distinguish between human-written and AI-generated text. Anthropic announced this summer it is implementing a new watermarking technology in newer versions of its Claude AI model.

The feature aims to comply with the EU AI Act and enhance transparency in digital communication, a topic relevant from local schools to community planning meetings. AI researcher Mustafa Ocal, a professor at FIU’s Knight Foundation School of Computing & Information Sciences, said the watermark is undetectable by humans but identifiable by a proprietary scanner. Ocal, who has analyzed Anthropic's approach, believes the development marks a significant step toward AI accountability.

“It’s brilliant,” Ocal stated. “People think, ‘I can just get rid of the watermark by deleting em dashes.’ But it’s much more complex than that. It is never, ever going to be human detectable.” The technique relies on subtle word choices dictated by a complex formula, making it difficult to remove without extensive editing.

While Google has used similar watermarking techniques since 2024, Anthropic's universal implementation across its latest models and its controlled scanner access for specific organizations like regulators and media outlets are key differentiators. Ocal explained the watermark is not based on specific words or sentence structures. Instead, it uses statistical patterns of word selection that deviate from typical human writing probabilities.

The watermark applies to text generated by Claude that is five sentences or longer. Access to the scanner is in a private preview and limited to organizations including law enforcement, regulators, media, and researchers. This controlled access aims to ensure the technology is used responsibly for verification purposes.