Claude’s Text Watermark: How It Works and What It Means
Claude Feature Guide
Since August 2, 2026, text generated by Claude carries an invisible, machine-readable watermark — worldwide, across every Claude product including the API. Here’s what it is, how it works, what it can and can’t prove, and what it means if you write with Claude.
Quick answer: Claude now embeds an imperceptible watermark in generated text and attaches C2PA provenance metadata to generated files. It doesn’t change the words you read, it survives copy-paste and light editing, and only Anthropic (and, soon, a detection API) can verify it. A watermark proves Claude processed the text — not that AI wrote it. There is no opt-out.
What Is Claude’s Text Watermark?
On August 11, 2026, Anthropic announced that Claude would watermark the text it generates, and on August 14 it published a technical explanation of how the system works. The trigger was regulatory: Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026, which implements the EU AI Act’s transparency requirements and, from August 2, 2026, requires AI providers serving the EU to mark AI-generated or AI-edited content in a way other systems can identify.
Anthropic chose to apply the marks worldwide, not just to EU users. Two mechanisms are involved:
- Text watermark — an imperceptible statistical pattern woven into the generated text itself. It travels with the text when copied and pasted.
- File provenance metadata — generated files such as SVG, PNG, and JPG images carry signed C2PA content credentials in their metadata.
How the Watermark Works
The watermark doesn’t insert hidden characters or change what Claude says. Instead, it changes the source of randomness Claude uses to choose between words when several options are equally good. When a model generates text, there are many low-stakes moments where two or three words would work equally well; the watermark uses a secret key to nudge those choices in a way that leaves a statistical fingerprint across the whole passage.
Technically, this happens at the token-sampling stage using an approach known as tournament sampling: candidate tokens are scored against key-derived signatures and the winner is selected. Because the effect only appears where alternatives are already equivalent, Anthropic says there is “no practical impact on the quality or content of Claude’s outputs” — watermarked and unwatermarked text are indistinguishable to a reader. Verifying the pattern requires the encoding key, which only Anthropic holds.
Where It Applies
| Scope | Details |
|---|---|
| Products | Claude (claude.ai apps), Claude Platform (API), Claude Code, Claude Cowork, and Claude Tag |
| Cloud partners | AWS, Google Cloud, and Microsoft Foundry deployments |
| Models | All models launched on or after August 2, 2026 at launch; existing models are being retrofitted |
| Geography | Worldwide — applied at the model level, so it’s present regardless of product, surface, or region |
| Text | Embedded watermark in generated text |
| Files | C2PA provenance metadata on supported file types (SVG, PNG, JPG) |
| Opt-out | None for users or developers |
The API point matters for developers: because the watermark is applied at the model level, text your application generates through the Claude API is watermarked too. See our Claude API guide and the current model lineup.
What the Watermark Does — and Doesn’t — Prove
This is the part most coverage gets wrong. Per Anthropic’s own documentation, a detected watermark “can only determine that Claude was likely involved with the content at some point. It cannot distinguish ‘Claude wrote this’ from ‘Claude heavily edited this.’” A human-written essay that you paste into Claude for proofreading, translation, or summarizing comes back watermarked. The mark proves processing, not authorship.
Other stated limits:
- Heavy editing removes it. Light edits leave the watermark intact; a complete rewrite, translation, or mixing with other writing can erase it. Anthropic has not published exactly how much editing is enough.
- Short passages are unreliable. The watermark is statistical, so a sentence or two doesn’t contain enough signal to detect.
- File metadata is fragile. Screenshots and re-saving strip C2PA credentials from files.
- Older models have no mark. Text from models that haven’t been retrofitted carries no watermark at all.
- Third parties can’t detect it yet. Detection needs Anthropic’s key. Anthropic says it “will soon be offering a watermark detection API” and is working to let users and third parties verify marks, with technical documentation to follow.
What It Means for You
- Students and writers: conventional AI detectors still can’t read the watermark — they never had the key. What changes is that Anthropic-backed detection may become available to platforms and institutions via API. Our Is Claude AI detectable? guide covers the broader detection picture.
- Content and SEO teams: the watermark has no effect on how the text reads, and Anthropic is not publishing the key, so it doesn’t identify Claude-assisted content to search engines today. It does mean “processed by Claude” is verifiable in principle. See Claude for SEO.
- Developers: API output is watermarked and there is no flag to disable it. If your product republishes Claude text, plan for the possibility that a detection API can later confirm Claude’s involvement.
- Privacy: the watermark encodes that Claude was involved — Anthropic has not described it as encoding user identity. For Claude’s wider privacy posture, see Is Claude AI safe?
Frequently Asked Questions
Does Claude watermark its text?
Yes. Since August 2, 2026, Claude embeds an imperceptible, machine-readable watermark in generated text and attaches C2PA provenance metadata to generated files. It applies worldwide across claude.ai, the API, Claude Code, Claude Cowork, and Claude Tag.
Can you see Claude’s watermark?
No. The watermark changes the source of randomness Claude uses to choose between equally good words, leaving a statistical pattern rather than visible characters. Anthropic says there is no practical impact on the quality or content of the text.
Can AI detectors detect Claude’s watermark?
Not currently. Verifying the watermark requires Anthropic’s secret key. Anthropic says it will soon offer a watermark detection API and is working to let users and third parties verify marks.
Does editing remove Claude’s watermark?
Light editing does not remove it, and the mark survives copy-and-paste. A complete rewrite, translation, or heavy mixing with other text can remove it. Short passages are also too small to detect reliably.
Does the watermark prove text was written by AI?
No. Anthropic states the watermark can only show Claude was likely involved at some point — it cannot distinguish text Claude wrote from human text that Claude proofread, translated, or summarized.
Is Claude API output watermarked?
Yes. The watermark is applied at the model level, so text generated through the Claude Platform API — and through AWS, Google Cloud, and Microsoft Foundry deployments — is watermarked, with no opt-out for developers.
Why did Anthropic add a watermark to Claude?
Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026. Under the EU AI Act’s transparency rules, providers must mark AI-generated or AI-edited content from August 2, 2026. Anthropic chose to apply the marks globally.