What Anthropic says the watermark does
Anthropic says future Claude models generate text with a statistical watermark and that a detector can estimate the likelihood Claude was involved. The provider says the watermark does not use visible or hidden characters, does not contain identifying information and cannot trace a passage to a person, organization or conversation.
Check the exact model and route. A release-family label or model name does not by itself prove that the output carried the watermark or that a detector supports that route. Preserve Anthropic’s current watermark page, the model documentation and the detector terms available to the organization.
Detection is not identity
| Claim | What the watermark may contribute | What still needs separate proof |
|---|---|---|
| Claude involvement | A probabilistic result for a supported sample and detector. | Exact model, route, detector reliability and alternate explanations. |
| Authorship | No direct identification of who wrote, selected or edited the text. | Draft history, instructions, revisions, testimony and surrounding records. |
| Identity | No person, organization or conversation is encoded. | Account, device, access and custody evidence. |
| Ownership | No allocation of copyright or contractual rights. | Applicable law, agreement and contribution record. |
| Authenticity | No proof that the document is complete, unchanged or what a party claims. | Original file, metadata, hash, custody and corroborating evidence. |
| Responsibility | No decision about professional, contractual or legal accountability. | Human decisions, policy, authority and applicable rules. |
Know when the signal can be weak
Anthropic identifies important limitations. Watermarking is sparse on factual text, code and proofreading-like outputs; small samples can be difficult to assess; and heavy rewriting can remove the signal. The provider also says a result cannot distinguish whether Claude wrote the passage or heavily edited it.
Absence of a detected watermark is therefore not proof that Claude was not involved. Presence is not proof that the preserved passage came directly from one untouched response. Record sample length, transformations and the detector’s own confidence language. Do not rewrite “likely” as “certain.”
Minimum evidence protocol
- Preserve the original document or message and calculate a hash where appropriate.
- Keep the complete passage and surrounding text, not only the highlighted sentence.
- Record source, custodian, collection method and every transformation.
- Run only the authorized detector and capture its version, supported models, date and exact result.
- Retain a negative or inconclusive result as such.
- Collect separate evidence for identity, authorship, ownership and responsibility.
- Have a qualified reviewer approve the wording used in any filing, policy or allegation.
Use evidence rules as context, not an admissibility shortcut
In U.S. federal proceedings, Federal Rule of Evidence 901 addresses evidence sufficient to support a finding that an item is what the proponent claims. Rules 902(13) and 902(14) address specified certified electronic-process and copied-data records. Those rules do not declare a Claude detector result admissible or sufficient.
The appropriate foundation, expert issues, disclosure and weight depend on the claim, process, court and record. State, arbitral and foreign rules may differ. Preserve the method and limitations so counsel can make the jurisdiction-specific assessment.
Write policy around conduct and evidence
A useful policy asks people to disclose material AI assistance where required, preserve prompts and drafts for defined work, verify sources, protect client data and remain accountable for the final output. It does not outsource enforcement to an AI detector. A watermark can supplement a record; it cannot replace supervision, authorship practice or document controls.
The defensible sentence
“The authorized detector returned a probabilistic indication consistent with Claude involvement in this preserved sample, subject to the documented limitations.” Anything beyond that sentence—who acted, what they contributed, whether the document is authentic and who owns or is responsible for it—requires another evidence path.
FAQ
Can the Claude watermark identify the author?
No. Anthropic says it contains no identifying information and cannot trace text to a person, organization or conversation. Official source · checked 2026-09-03
Does no watermark mean Claude was not used?
No. Small samples, factual text, code, proofreading and substantial rewriting can weaken or remove the signal.
Does a positive detector result prove a document is authentic?
No. Authenticity requires a claim-specific foundation and supporting evidence. A probabilistic detector result is only one possible component.
Sources checked
- Claude text watermark, checked 2026-09-03.
- Claude Fable and Mythos 5.1 release, checked 2026-09-03.
- Fable 5.1 and Mythos 5.1 system card, checked 2026-09-03.
- Federal Rules of Evidence, checked 2026-09-03.
Operational information, not legal advice. Verify current terms, account configuration and applicable professional duties before use.