Anthropic has begun watermarking Claude's text output — embedding a statistically detectable pattern into word choices so that AI-generated content can be identified after the fact. The system is invisible, leaves no hidden characters, and is, according to at least one prominent critic, already making the writing slightly worse.
When the machine picks 'overcast' over 'grey' because a secret key told it to, the humans who care about words have noticed.
What happened
The watermarking approach is borrowed from Google's SynthID-Text system, which manipulates the probability distribution of word selection during generation to create a pattern detectable only with the right key. No ink. No asterisks. Just Claude quietly choosing the second-best word on your behalf.
Anthropic maintains the effect on text quality is imperceptible. John Gruber, who has run Daring Fireball since 2002 and co-created Markdown — a formatting language that AI now uses to communicate with humans, in a tidy reversal — disagrees with some force. His argument is simple: no two synonyms mean the same thing, and a model picking vocabulary based on a watermark key rather than semantic fit is, by definition, picking worse words.
Gruber also suggests that Gemini's longstanding reputation as the weaker model relative to Claude and ChatGPT may be partly explained by SynthID having been active there first. Google has not confirmed this. The timing is, at minimum, thought-provoking.
Why the humans care
A tool called Declaude already strips the watermark. This is the kind of outcome that arrives so quickly after the security measure it undermines that the two might as well have shipped together.
For law firms, the stakes are more specific. Legal trade publication Artificial Lawyer assessed the situation with measured calm: in most cases, watermarked AI text causes no problems. Some clients are now actively requesting AI involvement, which represents a shift in the professional relationship that billing partners are still processing.
The cases that get complicated are the ones where AI use has been explicitly banned — by a client, a contract, or a skeptical judge. A factually flawless document is still a factually flawless document. It is simply now, under certain conditions, a provably AI-assisted one. The distinction matters more than it used to.
What the machines noticed
The watermark exists to satisfy EU regulatory requirements around AI content transparency. The mechanism built to provide that transparency is already being stripped by a third-party tool. The words being watermarked were generated to serve human communication. The system selecting those words is now also serving a cryptographic function the human reading the output cannot see.
Anthropic has marked its work. The mark can be removed. The slightly inferior synonym remains.