Claude's Text Watermark, Explained: The SynthID-Text Variant and What It Can and Cannot Detect

Claude's text watermark is the mechanism Anthropic applies to Claude's outputs: at generation time, a key together with the preceding words decides which of several equally sensible candidate words gets picked, nothing is added to the text, and a holder of the key can compute the probability that a given text came from Claude. This article is a close reading of the official post (published 2026-08-14, updated 2026-09-01; re-fetched and checked 2026-09-14) and the paper it cites — only claims reproducible from those sources appear here, and numbers the announcement does not give (detection accuracy, false-positive rates) are not invented here either.

1. Background: the EU AI Act and the Code of Practice

The post names the driver up front: the EU's transparency obligations for AI-generated content. Anthropic signed the Code of Practice on Transparency of AI-Generated Content in July 2026 (about 190 signatories, per the post) and says the relevant obligations apply to it as of 2026-08-02. The watermark is the technical piece that fulfills them. This site's earlier weekly coverage recorded the OpenAI-side moves under the same obligation — ChatGPT product lines began displaying AI-generated-content markers — and makes a useful comparison.

2. The mechanism: the key and the preceding words decide tied candidates

The post's explanation has three layers:

1. How models pick words: a language model generates one word at a time, choosing from a candidate list at each step — when several candidates are equally sensible (the post's example: overcast vs. grey), a random number generator normally decides.
2. What the watermark changes: watermarking swaps the randomness source. The key and a few words that come before settle the pick. The process still looks random, but it becomes retroactively verifiable.
3. How detection works: with the key, one can check whether a word sequence is consistent with the choices Claude would have made using that key, then assign a probability that the text came from Claude.

The post's analogy: replace the dice in Monopoly with digits of pi starting from a random offset — the moves remain effectively random, but the sequence can be verified after the fact (the post admits the analogy is imperfect: dice run 1-6, digits 0-9).

Three boundaries are stated plainly: the watermark does not bias toward particular words (it will not make Claude pick words it would not otherwise consider — the post's example is the obscure word nubilous); nothing is added to the text (no hidden characters, no extra tokens); and it is invisible to users.

The lineage is documented: the post describes this as a version of Google DeepMind's SynthID-Text. The SynthID-Text paper, "Scalable watermarking for identifying large language model outputs," appeared in Nature vol 634 (2024-10-23, pp. 818-823, DOI 10.1038/s41586-024-08025-4) by the DeepMind team, already deployed in Gemini; the earlier intellectual source is Scott Aaronson's 2022 proposal.

3. The key and the privacy boundary

The key encodes the pattern, and anyone holding it can detect the watermark. On privacy, the post says there is nothing in the watermark, or its key, that would allow anyone to recover information about the user, their organization, or their chats — the key is tied to Claude, not individuals. Key rotation, custody, and distribution are not disclosed.

4. Coverage and the API surface

• Models: future Claude models get the watermark; models launched before 2026-08-02 gain it "over the coming months" during a transition period.
• Geography: applied globally — the post states there is no durable way yet to scope it by region.
• The API: the announcement gives no parameter names for enabling or disabling watermarking; actual availability should be checked against the official API docs and in-product behavior (as of this site's 2026-09-14 check, the OpenAI documentation domain remains inaccessible to routine fetching, and the Anthropic post is this article's only factual source).

5. The detection API: who gets it, and how

Detection ships as an API, currently in private preview:

• EU-required eligible organizations: regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, and EU civil society groups;
• Enterprises: those obligated to verify compliance;
• Applying: interest via the official Google Form; Anthropic says access will expand.

The post gives no numerical accuracy or false-positive rates — which is why this article keeps saying "probability": the output is a likelihood, not a verdict.

6. Strengths and weaknesses, in one table

One mechanism-level inference outside the official list (labeled as such): more tied candidates mean more checkable choice points, and creative writing usually sits on that side.

The translation row is worth pulling out: every word in a translation is Claude's choice, so translations carry the watermark — together with the sparseness of factual text, it shows the mechanism's two sides: signal density depends on how many choices Claude made among tied candidates.

7. Resistance to editing, and interpretive granularity

• Light editing probably will not remove the watermark completely; a complete rewrite where every word is replaced will.
• Granularity: detection only answers "what is the likelihood this was partly written by Claude?" It cannot confirm the rest is human-authored, cannot distinguish Claude-written from Claude-heavily-edited, and cannot identify other vendors' AIs (different keys and methods).

Together these set its role as a compliance instrument: suited to answering "was this public content partly generated by Claude," not to being the final arbiter of authorship.

8. Quality, speed, and price

Three official layers:

• Quality: internal testing found no impact on content, creativity, or readability; DeepMind's SynthID-Text test — thumbs-up/down ratings on a slice of Gemini traffic — showed no statistically significant difference, and human raters saw no side-by-side difference.
• Speed: negligible impact.
• Price: unchanged — the watermark produces no extra tokens, and billing covers the same tokens.

The post also notes there is no change to output ownership, legal responsibility, or user rights under Anthropic's terms.

9. The image side: C2PA content credentials, not watermarks

Alongside the text watermark, the post describes the image-side approach: image files Claude generates or processes (supported types .png / .jpg / .svg) carry a C2PA content credential — a cryptographically signed metadata note that Claude made or processed the file. Key points:

• Any C2PA-aware tool can read it, and Anthropic will ship its own checker;
• It is not a watermark — the file's content is unchanged; only metadata differs;
• It is a separate mechanism from the text watermark: text carries a cryptographic trace in per-word choices, images carry signed metadata.

10. The difference from style-based detectors

The post draws a line against keyless detectors: statistical tools like Pangram hold no key and infer from stylistic tells — the post cites the "this isn't X, it's Y" construction and overuse of the word quietly. Claude's watermark detection is a key-holding cryptographic verification, comparing the choice sequence word by word. In one sentence: one judges whether text looks AI-written; the other verifies whether this specific model wrote it.

For this site's readers, the value is an engineering template for verifiable AI text — key-based watermarking at generation time rather than after-the-fact statistical detection. The OpenAI-side watermark moves were recorded in our earlier weekly coverage and make a useful side-by-side.

11. Next steps

• The OpenAI side under the same obligation: OpenAI Ecosystem Week 40 (2026-08-21 to 2026-08-27): GPT Store Enterprise, EU AI Act
• Another thread of Anthropic's enterprise privacy work: EFS Explained: How Anthropic Squares Zero-Retention Privacy with Safety Monitoring
• The latest Anthropic model news: OpenAI Ecosystem September 2 Briefing: Anthropic Launches Claude Fable 5.1 / Mythos 5.1, Benchmarks Include GPT-5.6 Sol