Claude's new watermark is basically vibes-based word roulette
Anthropic finally explained how Claude will secretly mark its text as AI-written, and it's not invisible characters or sneaky metadata. Instead, whenever the model faces a 'low-stakes' word choice, like picking 'grey' versus 'overcast,' it uses a secret key to bias which word gets picked, leaving a statistical fingerprint only Anthropic can detect. The catch: treating synonyms as totally interchangeable ignores that writers pick specific words on purpose, which is exactly the point critics say AI still doesn't get about writing.
Nothing says 'we respect the craft of writing' like reducing it to a coin flip with extra steps.