Anthropic Will Watermark Claude by Changing Word Choice

Anthropic detailed how Claude's text watermarking will work: rather than adding hidden characters or metadata, it biases the random selection among near-equivalent words in a pattern only Anthropic's detector can read.

Trace Cohen
By Trace Cohen · Value Add Pulse AI Desk
Founder & Editor-in-chief, Value Add VC
3 min read
Anthropic Will Watermark Claude by Changing Word Choice: key figures
ShareXLinkedInEmail

THE RUNDOWN

1

Anthropic said in a weekend blog post that 'nothing is added to the text and there are no hidden characters,' and that watermarked and unwatermarked text will be indistinguishable to readers, per 404 Media

2

The method alters 'the source of the randomness used to pick among words,' letting Anthropic build a detector for Claude-generated text

3

Anthropic's own example -- 'overcast' versus 'grey' after 'The weather today was cold and...' -- assumes those choices are interchangeable to a reader

4

TechCrunch separately reported Claude users objecting that the watermarks will expose their use of the tool at work and in classrooms

The VC Read

Value Add VC analysis

Watermarking that dies to a paraphrase is a compliance artifact, not a detection system, and Anthropic almost certainly knows that -- the EU AI Act needs a provenance mechanism to exist, not to be robust. The commercial angle founders should notice: if labs bias word selection for watermarking, any product whose value depends on exact model output distribution now has a vendor-controlled variable it cannot see. Ask your API provider whether watermarking is on by default and whether you can opt out.

Analysis

Anthropic announced earlier this month that future versions of Claude would watermark their output. This weekend it explained the mechanism, and the answer is more interesting than hidden characters: Anthropic will change how Claude writes.

"Nothing is added to the text and there are no hidden characters," the company wrote in its blog post, quoted by 404 Media. "The difference between watermarked and un-watermarked text will not be distinguishable to readers." The technique modifies "the source of the randomness used to pick among words" -- a pattern known only to Anthropic, which can therefore build a detector that reads it back out.

The company's own illustration is where the criticism lands. Take "The weather today was cold and..." -- the next word is unlikely to be "sugary" and quite likely "overcast" or "grey." Anthropic argues that in cases like this "it doesn't matter much to the reader which of these latter two words the model ultimately chooses," so the choice can be settled by a controlled random number instead of a free one. 404 Media's Jason Koebler takes that as evidence AI companies treat words as interchangeable tokens and have no framework for judging writing quality -- "overcast" and "grey" carry different registers, and the space of near-equivalent words is exactly where a writer's voice lives.

“The company's own illustration is where the criticism lands.”

Fragile by Design

The engineering is sound as far as it goes, and it is the same family of approach Google demonstrated with SynthID-Text. It is also fragile in the obvious way: paraphrasing through a second model, heavy human editing, or round-tripping through translation degrades the statistical signal. A detector that works on unedited output and fails on lightly rewritten output is a tool for catching careless users, not determined ones.

The competitive dynamics of watermarking are more interesting than the technique. If Anthropic watermarks and OpenAI does not, Claude output becomes detectable while GPT output does not -- which, for any user whose motive is to avoid detection, is a direct reason to switch providers. That is a real commercial cost to being first, and it is why voluntary provenance schemes tend to require either regulatory mandate or industry-wide coordination to survive. The EU AI Act's transparency obligations supply the mandate in Europe; nothing equivalent applies in the United States.

Detection accuracy is the number Anthropic has not published. A watermark detector has two error rates that matter in opposite directions: false positives accuse a human writer of using AI, and false negatives let AI text pass. Universities and employers will use whatever tool exists regardless of its stated confidence interval, which is what happened with the first generation of AI-detection products and produced a documented pattern of wrongly accused students, disproportionately non-native English writers. Anthropic controls both the watermark and the detector here, so it is the only party that can publish those rates -- and until it does, institutions deploying the detector are making accusations they cannot calibrate.

That is precisely who is objecting. TechCrunch reported that some Claude users are upset the watermarks will catch them using the tool at their jobs and in classes -- a complaint that inadvertently confirms the feature works on the population it was built for. Pulse has covered Anthropic's watermarking plans as EU AI Act transparency obligations came into force, and the compliance driver is unchanged: labs need provenance signals regulators can point to, whether or not the signal survives contact with a paraphraser.

ShareXLinkedInEmail

More from the Pulse archive: Artificial Intelligence coverage → · All topics

Key Sources

1 source

Reported by 404 Media · Analysis by Value Add Pulse.

← Back to Pulse

THE WIRE in your inbox— Tech, startup & VC news with The VC Read, a few times a week. Free to subscribe, no spam.

No spam, unsubscribe anytime. Sent via Substack · Privacy · or subscribe on Substack ↗