Anthropic Embeds Watermarks in Claude AI Output

Anthropic has started embedding a statistical watermark into everything Claude writes, a method built on Google's SynthID-Text approach that nudges which words the model picks during generation without inserting hidden characters or visible tags. The company says this satisfies EU transparency rules requiring AI content to be identifiable, and insists the practice leaves creativity, readability, and meaning untouched. Some clients have specifically asked for the feature, according to Anthropic, treating it as a compliance box to check rather than a burden.

Not everyone buys the "no effect" claim. The watermark works by shifting the probability of certain synonyms over others based on a hidden key, which means Claude sometimes favors a word not because it fits best but because the key says so. Tools already exist, like Declaude, that can strip the marking out entirely, raising the question of what the watermark actually accomplishes if it's this easy to defeat.

For law firms in particular, the stakes go beyond writing style. Permanently detectable AI text could surface at inconvenient moments, from fee disputes to contracts that explicitly bar AI-generated work.