Your Claude Text Carries a Watermark You Cannot Turn Off or Check. Grok's API Watermark Starts November 7, and Admins Can Opt Out.
Anthropic's help center says it signed the EU AI Act Article 50(2) Code of Practice and that supported Claude models embed machine-readable text watermarks across the API, the apps, Claude Code, Claude Cowork, Claude Tag, and the cloud partners, worldwide. Detection is a private preview for eligible organizations, and there is no user opt-out. Anthropic's own limitations note says Opus 4.6 and Sonnet 4.6 may not carry the mark reliably under non-default temperature, top_p, or top_k. xAI takes the opposite posture: from November 7, listed Grok models watermark by default on every API endpoint, with an x-watermark response header and a per-model opt-out for team admins in the console. We read both help pages.
By T.W. Ghost
Two labs now watermark the text their models produce, and they have chosen opposite postures. Anthropic's help center says the company signed the EU AI Act Article 50(2) Code of Practice on Transparency of AI-Generated Content, and that supported Claude models embed machine-readable text watermarks across the Claude Platform API, the Claude apps, Claude Code, Claude Cowork, Claude Tag, and the cloud partners, AWS, Google Cloud, and Microsoft Foundry, worldwide. Detection is a private preview for eligible organizations: regulators, law enforcement, media and fact-checkers, researchers, educational organizations, EU civil society, and enterprises with obligations under the Act. There is no user opt-out, and as a user you cannot run the detector yourself.
xAI's developer docs say that from November 7, 2026, the listed Grok models watermark by default on every API endpoint, covering text, structured outputs, and tool-call arguments; responses carry an x-watermark: true header; and a team admin can opt out per model in the console, with the change taking effect in about five minutes and landing in an audit log. Same regulatory driver, two designs: Claude's mark is on, invisible to you, and permanent; Grok's is on, visible in a header, and switchable by an admin. We read Anthropic's help article and xAI's watermarking page, and The Decoder's report of a customer email for the rollout timing. Here is what held up.
Two Watermarks, Side by Side
| Claude (Anthropic) | Grok API (xAI) | |
|---|---|---|
| Basis | EU AI Act Article 50(2) Code of Practice, applied worldwide | Watermarking docs; default on from November 7, 2026 |
| Surfaces | API, Claude apps, Claude Code, Claude Cowork, Claude Tag, AWS, Google Cloud, Microsoft Foundry | Every API endpoint for listed models |
| Models | Help-center table includes Fable 5.1 and 5, Mythos 5.1 and 5, Opus 5.5, 4.8, 4.7, 4.6, 4.5, Sonnet 5.5, 4.6, 4.5, Haiku 5.5 and 4.5; cloud-partner columns per table, with a Bedrock rollout footnote for some | grok-4.7, 4.6, 4.5, 4.3, grok-4.20-reasoning, grok-4.20-non-reasoning, grok-4.20-multi-agent |
| Coverage | Text | Text, structured outputs, tool-call arguments |
| Visible signal | None to the user | x-watermark: true response header |
| Opt-out | None | Team admin, per model, in the console; about five minutes; audit-logged |
| Detection | Private preview for eligible organizations | Not the subject of the page we read |
| Known gap | Opus 4.6 and Sonnet 4.6 may lack a detectable mark under non-default temperature, top_p, or top_k (Anthropic's limitations) | Not stated |
You cannot promise unmarked Claude output. If you build on a supported model and tell a customer the text is clean, Anthropic's help page says otherwise, and you have no detector to prove either way.
The sampling loophole is Anthropic's own text. The limitations section, not a press report, says Opus 4.6 and Sonnet 4.6 may not reliably carry the watermark when sampling parameters are non-default. Two older models, one documented gap.
xAI's design is the one an admin can reason about. A header you can log and a switch you can flip are auditable. Whether to flip it is a policy decision; that you can is the difference.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Anthropic signed the Article 50(2) Code of Practice; marks apply worldwide on supported models and surfaces | Verified (Anthropic help center) |
| Detection private preview; eligible-organization list; no user opt-out | Verified (help center; opt-out verified as absent) |
| Model coverage table | Verified (help center table) |
| Opus 4.6 and Sonnet 4.6 gap under non-default sampling | Verified (help center limitations) |
| xAI default-on from Nov 7; model list; text, structured outputs, tool-call arguments | Verified (xAI docs) |
x-watermark header; admin opt-out per model; about five minutes; audit log | Verified (xAI docs) |
| Customer email naming a November 3 rollout for specific older models | Press (The Decoder) |
| OpenAI's equivalent defaults | Unverified (no OpenAI primary read this pass) |
What To Do
Builders on Claude: assume output from supported models is marked, say so in your own terms, and do not promise otherwise.
xAI teams: decide the opt-out policy before November 7, make the change in the console, and keep the audit-log entry.
The frontier is elsewhere: the marks say nothing about capability, and the same month's other transparency move, the shared robots.txt no-training token, is also a lab-by-lab story. Our model comparison has the current standings side by side.
Sources
- Claude support: how Claude marks AI-generated content
- xAI docs: watermarking
- The Decoder: Anthropic's watermark rollout leaves a loophole on two older Claude models (press, for the email timing)
*Not sure which lab's output policy fits what you promise your own customers? Take the free 2-minute quiz and get matched. Then read the other default that changed this month on the data side: Gemini API logging for new paid projects.*