Claude to add invisible watermarks to AI text under EU rules, changing how it picks words

nordicinst · x · 2026-08-18

Anthropic said over the weekend that Claude will change how it generates prose at a granular, random level to comply with an EU regulation requiring all AI-generated text to be watermarked from December. The watermark leaves a pattern detectable by Anthropic and anyone able to decode it, but invisible to the average reader.

The move sparked debate over writing quality: veteran tech blogger John Gruber called it "a perverse adulteration of what it means to write," arguing the constraint will force Claude into worse, less precise word choices overall, even if accuracy doesn't suffer. The Guardian also notes familiar LLM prose tics—overused "delve", em dashes, "it's not X but Y" constructions—as context.

Original post →

More from Models

Models channel →