Industry analysis · August 11, 2026
Claude text watermark: what Anthropic plans and what it proves
Claude watermark detection is in private preview. Learn who can request access, what a result means, and how to keep evidence for an AI-assisted draft.

Anthropic plans to add a SynthID-Text watermark to future Claude models. A detector will estimate whether Claude helped produce a passage, but it cannot prove who wrote it, how much a human changed, or whether any rule was broken.
Sources checked October 26, 2026. This refresh corrects detector availability using Anthropic’s September 1 update and adds an editorial review exercise. We have not accessed the private detection API or measured its accuracy.
Anthropic announced on August 14, 2026 that future Claude models will generate watermarked text. Its public wording is forward-looking. The announcement does not say that every current Claude response already carries the mark.
The planned system uses Google’s SynthID-Text method. It changes the statistical pattern of token choices while preserving the visible answer. There are no hidden characters, no file metadata to remove, and no user, organization, or conversation identifier inside the watermark.
Can you check a Claude watermark today?
Anthropic’s updated announcement describes access for eligible organizations, including media, researchers and educators, plus enterprises with relevant verification obligations. Follow its registration link to request access; registration does not establish approval. The page does not offer an unrestricted public text checker or a model-by-model rollout list.
Google DeepMind’s open-source SynthID Text implementation is a research reference. Its detector requires the same watermark configuration used during generation, including the key. Running the sample notebook with your own key therefore does not establish whether a passage carries Claude’s watermark. That is a technical inference from the matching-configuration requirement, not a test we performed.
What the detector can tell you
Anthropic says its detector will estimate how likely it is that Claude wrote at least part of a passage. That can support a provenance review. It cannot answer these questions on its own:
- Did a human write the first draft and use Claude only for editing?
- What percentage came from Claude?
- Did the writer violate a school, employer, client, or platform policy?
- Did another AI system write the unmarked parts?
- Is the passage true, original, or well sourced?
A positive result is evidence that Claude may have contributed to the text. It is not proof of sole authorship or misconduct.
Where detection gets weaker
The signal needs enough model-generated text to become statistically visible. Anthropic warns that detection is less reliable for short passages and constrained outputs such as factual answers, code, or proofreading. Heavy editing can also weaken it. A complete rewrite can remove it.
That limitation is expected. The watermark lives in word-selection patterns, so replacing enough words changes the pattern. It also means a negative result does not prove that no AI was used.
Google DeepMind’s SynthID explanation says detection works best on longer responses and becomes weaker after translation of generated text or on factual answers. Its 2024 research paper evaluates the underlying technique; it does not establish the accuracy of Anthropic’s private API on your documents.
| Result | Safe conclusion | Unsafe conclusion |
|---|---|---|
| Strong positive | Claude likely generated part of the passage | Claude wrote all of it |
| Weak or uncertain | The detector does not have enough evidence | A human wrote it without AI help |
| Negative | The Claude watermark was not detected | No model generated or edited the text |
These are review categories for this article, not documented API response labels. Interpret any actual response using the documentation supplied with your access.
What Anthropic says will not change
The company says watermarking will not add tokens, visible wording, or extra user cost. It also says the signal does not encode personal or account information. Older Claude models will move to watermarking later rather than changing immediately.
Anthropic links the launch to transparency requirements under the EU AI Act and its Code of Practice. The first rollout is planned globally, not only for European users.
How teams should use it
- Publish an AI-use policy. Define whether drafting, editing, translation, or summarization must be disclosed.
- Keep source and revision evidence. Draft history and review notes explain mixed authorship better than one detector score.
- Use detection as one input. Ask the writer what tools were used and review the work itself before making a consequential decision.
- Test the detector on your own workflow. Short emails, code, and heavily edited copy may behave differently from long, unchanged model output.
- Keep a human accountable for the final work. A watermark does not check facts, citations, permissions, or quality.
The practical change is narrower than the headline. Anthropic is adding a way to detect likely Claude involvement in some text. It is not adding a universal AI detector or an authorship certificate.
Try it: prepare a draft-review worksheet
You can complete this exercise without detector access. You need a text editor and a folder for non-sensitive practice documents. Claude access is optional until you create the AI-assisted samples. The deliverable is a traceable set of drafts, not an accuracy score.
- Write a human baseline. Draft a few paragraphs about a fictional community garden without AI assistance. Save the original as
human-original.txtbefore any edits. - Save a generated draft separately. If you have Claude access, ask it for a new explanation of that fictional garden. Save the exact prompt and response, generation date, product and model label shown. Mark watermark coverage as unknown unless you have documentation for that model. If you lack access, leave this sample pending.
- Create two editing cases. Ask Claude to proofread a copy of your human baseline, keeping its meaning and wording where possible. Separately, edit a copy of the generated draft yourself. Preserve all originals and record which changes you made. These cases represent different directions of assistance.
- Complete one record per sample. Use the template below. Keep the detector field at
not rununless an authorized detector actually returns a result. - Review the work. Compare each final text with its original, verify any factual claims, and write one sentence explaining the tool’s role. Have another reviewer check whether the record supports that sentence.
Sample ID and original file:
Known origin: human / generated / mixed
Product, model label, and generation date:
Prompt and unchanged response file:
Edit history and final file:
Watermark coverage evidence: unknown / source and date
Detector access and version: unavailable / recorded details
Detector result: not run / exact returned result
Source checks and reviewer:
AI-use disclosure for this draft:
Unresolved questions:
Completion check: every final file has a preserved original, an origin record and an honest detection status. Another reviewer should be able to distinguish a human draft proofread by Claude from a Claude draft edited by a human without guessing from writing style.
If your organization receives preview access, use the provider’s current instructions and permitted test material. Save the returned result alongside the submitted sample and detector version. Compare it with the known origin; record disagreements rather than rewriting the origin label to match the detector. A small practice set helps inspect a workflow but cannot establish a production error rate.
Recover when the evidence is incomplete
| Problem | Next step |
|---|---|
| No detector access | Finish the provenance worksheet; leave detection as not run |
| Model or watermark coverage is unknown | Keep it unknown and retain the visible model label and generation date |
| A result disagrees with known draft history | Preserve both records and seek review; do not infer misconduct |
| Only a short excerpt is available | Record the missing context and request the original through an approved channel |
| An agent has replaced the original draft | Recover version history if available; otherwise record the evidence gap |
For an agent-assisted publishing workflow, ask the agent to produce the worksheet beside the draft and require a reviewer to verify both before publication. The model deployment decision lesson helps document where drafts and logs go. The Claude Code task-extraction tutorial offers a related exercise in retaining evidence and checking structured output.
Common questions
Can a hidden-character scanner detect this watermark?
No. Anthropic describes a statistical text signal, so a Unicode or metadata scan is not a test for it. File content credentials are a separate mechanism. Anthropic’s explanation.
Should I report a detector score as a percentage of AI-written words?
No. Keep the returned value and its documented meaning together. Do not turn a statistical result into a word count, an authorship percentage or a policy judgment.
Sources
Checked October 26, 2026. Provider statements above are attributed claims; the worksheet is an authored exercise, not a reported detector trial.
- Anthropic: How Claude’s text watermark works — published August 14, 2026; updated September 1, 2026 with private-preview access information.
- Google DeepMind: SynthID — official overview of text and other media watermarking.
- Google DeepMind: Watermarking AI-generated text and video with SynthID — text detection limitations.
- Google DeepMind: SynthID Text repository — reference implementation and matching watermark configuration requirement.
- Nature: Scalable watermarking for identifying large language model outputs — research paper published October 23, 2024; evidence about the method, not a Claude API evaluation.
Put this to work
Separate provenance evidence from authorship, intent, and misconduct.
Try
Build a small sample folder with original text, edit history, model details, and a review worksheet.
Prove it worked
Verify that every sample has a provenance record; record detection as not run unless you have approved access and an actual result.
Where it can pay
Teams will need editors who can review mixed-authorship work without treating a detector score as a verdict.
Keep in view
- Anthropic's September 1 update describes a detection API in private preview for eligible organizations, with broader access planned.
- Anthropic says future Claude models will add a SynthID-Text watermark; it has not announced that every current Claude response is already marked.
- The watermark changes token selection, not visible text, metadata, hidden characters, or user-identifying information.
- Detection estimates likely Claude involvement and becomes less reliable after short, factual, heavily edited, or rewritten text.