Anthropic's watermarking treats synonyms as fungible—sparking a debate on authorship
anthropic launched a text watermarking method that alters word choice in low-stakes decisions. critics argue this devalues writing by treating nuanced synonyms as interchangeable variables.
anthropic revealed a watermarking system for future claude releases that modifies the model's random seed during word selection to create a mathematically detectable pattern. the company uses examples like swapping 'overcast' for 'grey'—instances where anthropic claims the distinction is low-stakes—to embed the signal without impacting text quality or metadata. this approach has drawn criticism from writers and academics who argue that specific word choices are rarely interchangeable and that the technology reduces language to mere probability. the backlash highlights a growing tension between ai safety research goals and the nuances of human communication.