Anthropic's ‘watermark’ text adulteration in Claude is a perversion of writing
TL;DR Highlight
Anthropic will watermark Claude's word choices, and Gruber says this degrades writing quality.
Who Should Read
Backend developers shipping Claude output to users who want to understand how text watermarking affects quality and verification.
Core Mechanics
- The watermark is word choice, not hidden characters
- Gruber first guessed invisible Unicode characters.
- It actually makes the model slightly more likely to pick 'green' tokens at each step.
- A secret key splits green and red lists
- The lists are set deterministically on the fly at each token.
- Only the key holder can detect it, and there is no fixed list of favored words.
- Longer text gives more confident detection
- It works like flipping a coin more times to judge bias.
- It applies to text over 200 tokens (~150 words).
- Only Anthropic can detect Claude's mark
- Keys are held by each provider.
- Claude cannot read Gemini's watermark and vice versa.
- Gruber's core objection is that synonyms differ
- He says 'leaped at the chance' and 'jumped at the opportunity' are not the same.
Community Reactions
- Opinion that LLMs already sample randomly
- The watermark just replaces the random number generator with a seeded pseudo-random one.
- Others said T=0 isn't best writing.
- Concern that detection means sending text to every provider
- Universities, lawyers and editors would send unpublished text to many companies.
- Some want a shared public or self-hostable detection oracle.
- Info that the EU rules require interoperability
- A quoted clause requires a standard API, public signpost, or consortium detection by February 2, 2027.
- Explanation that a non-distorting variant keeps quality
- Commenters said it doesn't alter the underlying distribution but is weaker on short text.
- Another commenter worried it still shifts learned probabilities.
How to Apply
- Find which outputs exceed 200 tokens
- Pull Claude response lengths from your logs today.
- List the features likely affected.
- Audit AI-detection paths for external text transfer
- Check whether detection sends text to third-party APIs.
- Set consent or masking rules for unpublished documents.
- Log rewrite steps in your pipeline
- Mark translation, summarization and rewrite stages.
- Separate them from raw model output when reading detection results.
Terminology
Related Papers
The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge
Warning an LLM judge about bias or letting it self-critique improved 20 of 21 model–bias combinations.
90% of US working hours are spent on tasks not covered well by AI benchmarks
Mapping 29 benchmarks to job tasks shows 90% of US working hours lack good coverage
Postgres 19 delayed after AI reviewers find more bugs than humans can fix
LLM code review surfaced so many bugs that Postgres 19 slipped to November and 8-10 features were reverted
Show HN: Gemini 4 Argon plays StarCraft
LLMs wrote StarCraft bots in one hour; GPT-6 Astra scored highest.
A Plausible Explanation Is Enough to Mislead Multimodal Judges
Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.
Neural networks are clean
Looking at neurons directly shows most have one main job, challenging superposition.