# Claude's text watermark: what caching and editing change

URL: https://infro.io/blog/claude-s-text-watermark-what-caching-and-editing-change
Author: INFRO Editorial (https://infro.io/blog/editorial)
Publisher: INFRO
Published: 2026-11-08
Last modified: 2026-11-08
Last reviewed: 2026-11-08 (next review due 2026-12-08)
Topic: Security and governance
Format: Brief

> The watermark lives in word choices, so verbatim copies keep it, while truncation, light edits and non-Claude fallbacks dilute or remove it.[^claude-watermark] Treat it as probabilistic evidence of Claude involvement, never as proof of authorship, and record which model and which edits touched each stored text.

Key takeaways:
- The watermark is a pattern in which words Claude picks, not added characters, so an exact copy of a response keeps it.
- Truncation, light edits and proofreading shrink the evidence; a full rewrite removes it, and detection is weak on short samples.
- A translation by Claude is watermarked, but a summary or rewrite by a different model carries that model's marking, not Claude's.
- Log the serving model and every post-processing step next to stored text, so a later detection result can be read correctly.
- A watermark says Claude was probably involved; it says nothing about ownership, authorship or who sent the request.

## What the watermark is, and what it is not

Anthropic says future Claude models will generate text that contains a watermark, a way to estimate the likelihood that Claude was involved in writing it. It says it is doing this to comply with the EU AI Act.[1] The marking obligations entered into application on August 2, 2026, and about 190 organisations, Anthropic among them, signed the Code of Practice on Transparency of AI-generated Content.[2]

The mechanism matters for a gateway. Nothing is added to the text and there are no hidden characters. The method is a version of SynthID-Text: it changes the source of the randomness used to pick among near-equivalent words, using a key and a few preceding words.[1] So the signal is a statistical pattern across many word choices. It survives whatever preserves those words, and weakens as they are removed or replaced.

Limits Anthropic states: detection works poorly on small samples, confidence grows with length, factual passages and code carry less watermark, and the check only answers "how likely was Claude involved". It cannot say a text is human-written or that another AI wrote it.[1]

## Which gateway operations keep, dilute or drop it

The "Effect" column follows from Anthropic's description of the mechanism. The reading for stored copies is our inference, not a measured result.

| Operation | Effect on the watermark | Note for reviewers |
| --- | --- | --- |
| Exact-match cache replay | Kept: the stored text is the original answer | Same text served again, possibly to many users |
| Truncation or length cutoff | Diluted: fewer word choices left to test | Short fragments may not be detectable |
| Light edit or proofreading | Mostly kept in long text, weak in short | Detection can be too weak to register[1] |
| Full rewrite of every word | Removed[1] | Arguably no longer AI-generated text |
| Translation by Claude | Kept: every word is Claude's choice[1] | Translating with another tool replaces it |
| Summary by Claude | Present, since Claude chooses the words | Short summaries give little to detect |
| Fallback to a non-Claude model | Not Claude's watermark; that model has its own key or method[1] | Record which model answered |
| Code output | Less watermark; comments may carry some[1] | Do not rely on it for code provenance |

Two further cases. Concatenating Claude text with human or other-model text yields a mixed document, and the watermark can only attest to the Claude-chosen portion. Redaction by a guardrail changes words in place; the result is edited Claude text, not the original.

## Caching, logging and fallback in practice

An exact-match cache returns the stored answer verbatim, so the watermark is preserved on every hit. INFRO's response cache is opt-in and exact-match, and it only stores a response when output guardrails left it untouched and every choice finished with stop or tool_calls.[3] That means cached entries are original provider text, not edited text, and a truncated answer (finish reason length) is never replayed.[3] If you build your own cache, decide explicitly whether it may store post-processed text, and label which kind each entry is.

Logging can truncate. When INFRO records a request body, `content_truncated` is true if the stored copy was shortened, and a long string keeps its beginning and end around a marker naming how many characters were cut.[4] If you plan to run detection on archived answers, that logged copy is not the original text. Payload logging is off by default and retention is yours to choose, so confirm what you actually keep before promising anything.

Fallback changes the model. Connection failover is same-model, but model fallbacks are cross-model and opt-in, and the response names the model, provider and connection that served the request.[5] Store that field beside the text. A reviewer who later gets a negative Claude detection result needs to know the answer came from a different model, not from an unmarked Claude.

## What to tell a reviewer

- The gateway does not add, remove or verify watermarks. It passes through whatever the provider returns, plus any edits you configure.
- Verbatim copies (cache hits) retain the pattern; truncated, redacted, merged or rewritten copies may not.
- A detection result is probabilistic. A positive means Claude was likely involved at some point; it cannot separate writing from heavy editing, and a negative does not prove human authorship.[1]
- Detection access is limited. Anthropic describes a private-preview detection API for eligible organizations and for enterprises obliged to verify watermarking for their own compliance; others can register interest.[1]
- Ownership and legal responsibility do not change because of a watermark.[1]

Open questions to put to Anthropic or your cloud provider: whether older Claude models you still call are marked yet (Anthropic describes a transition period for models launched before August 2, 2026), and whether Claude served through a cloud platform carries the same marking. Anthropic's page does not answer the second.[1]

> **Note: Where INFRO helps**
> The response cache, request traces and per-request model attribution record which model answered, whether a response came from the cache and whether a logged copy was shortened. They do not detect watermarks.

## Questions

### Does the watermark change token counts or price?

Anthropic says watermarking produces no extra tokens and the model costs the same to serve and use, so estimated cost per request should not move because of it.[1]

### Can a watermark be traced to our organization or a user?

Anthropic states the watermark and its key carry nothing that would let anyone recover information about the user, their organization or their chats.[1]

## Sources

1. How Claude's text watermark works — Anthropic. https://www.anthropic.com/news/claude-text-watermark (accessed 2026-11-08)
2. Strong backing for the Code of Practice on Transparency of AI-generated Content — European Commission. https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content (accessed 2026-11-08)
3. Response cache — INFRO. https://infro.io/docs/features/caching (accessed 2026-11-08)
4. Request traces — INFRO. https://infro.io/docs/observability/traces (accessed 2026-11-08)
5. Failover & fallbacks — INFRO. https://infro.io/docs/features/fallbacks (accessed 2026-11-08)
