đź•’ Created · Updated

Anthropic to Implement Invisible Watermarks on Claude AI Models to Comply with European Union Regulations

Anthropic is introducing an invisible watermark to its Claude AI models to comply with the European Union's AI Act. The administration announced that the watermark will be applied to text and images generated by the AI, allowing users to identify the likelihood that Claude was involved in the content. Unlike physical watermarks, these marks are indistinguishable to the human eye and do not impact the quality, speed, or cost of the output. To achieve this, Anthropic uses a cryptographic key to influence low-stakes word choices during the generation process. While the watermark is effective for long-form text, it may be less reliable for short passages or highly factual statements where word choices are limited. The move has sparked a debate among users regarding credit for AI-generated work. While some users expressed frustration that their proofread text would be marked as AI-generated, others noted that the watermark provides a necessary level of transparency for the era of AI content.

Sources