Grok's Invisible Text Watermark Goes Live Today — and Claude's Next Wave Lands 16 November

 

Grok's Invisible Watermark Goes Live Today

From today, 7 November 2026, text generated by seven Grok models through the xAI API carries an invisible watermark by default, to comply with the EU AI Act. You will not see it, and the only trace in a single response is a new header, x-watermark, set to true. A text watermark is a statistical mark woven into the words a model chooses, invisible to the reader, that lets the provider later test whether a piece of text came from its model. The change covers every text generation API and every endpoint. Team admins can switch it off per model, but only after confirming that a legal exception applies. It covers the developer API only.

Article 50 of the EU AI Act requires providers of AI systems that generate text, audio, images or video to mark their outputs in a machine-readable format, detectable as artificially generated. That rule has applied since 2 August 2026, and systems already on the market before that date have until 2 December 2026 to comply. 7 November falls 25 days before that deadline. Anthropic has been rolling the same mark across Claude.

The Seven Grok Models, and What Gets Marked

Model	Watermark from 7 Nov 2026
grok-4.7	On by default
grok-4.6	On by default
grok-4.5	On by default
grok-4.3	On by default
grok-4.20-reasoning	On by default
grok-4.20-non-reasoning	On by default
grok-4.20-multi-agent	On by default

The watermark is enabled on every SpaceXSI API endpoint, including the global endpoint at api.x.ai and any regional endpoint, and it applies to every text generation API, including the Responses API and the Chat Completions API. It covers the generated text in the response, including structured outputs and tool call arguments — which matters for agent builders.

Illustration created for INORYVA

How the Watermark Works, and How to Check It

The mark adds nothing visible to the text. While the model generates, it slightly shapes which tokens it picks, so it adds no characters, markup or hidden text. There is no request parameter to send, and the response format is unchanged apart from the x-watermark header. To check a response, read that header: a value of true means that request's output was watermarked, and any other value, or no header at all, means it was not.

Detection itself is statistical. SpaceXSI scores a piece of text against the watermark and reports whether it is watermarked, not watermarked, or uncertain, and longer text gives a stronger signal. But detection access is granted by application, so today no journalist, teacher or company can check a suspicious Grok text on their own. xAI states that watermarking does not change response quality or speed — its own claim, with no independent testing behind it.

How to Opt Out — and What Opting Out Means

Only team admins can turn the watermark off, in the SpaceXSI Console under Team settings. The setting is per model for the whole team and applies to every API key on it. To opt out, an admin clears the checkbox for each model and confirms with "Opt Out". Changes apply to new requests within 5 minutes. Some models may be locked on.

Opting out is a legal decision, not a technical one: the docs ask you to confirm that an exception under applicable law applies, and that you are solely responsible for any marking or disclosure. If you serve EU users and cannot name a legal exception, leave the watermark on.

Illustration created for INORYVA

Claude's Parallel Rollout — and a Corrected Timeline

Anthropic is doing the same thing to Claude, with less choice. Every Claude model released since 2 August 2026 shipped with an invisible watermark on day one. Older models are getting the mark in waves. The first wave went live on 3 November 2026, covering Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6. A second wave follows on 16 November 2026, covering Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.5 and Claude Haiku 4.5. Anthropic emailed API customers about the second wave on 4 November (The Decoder). Opus 4.6 was due on 3 November and was pushed back two weeks without a reason given.

Two differences matter. First, Claude customers cannot turn the mark off — there is no opt-out, unlike Grok's per-model switch. Second, on older Claude models the watermark only applies when temperature, top_p and top_k stay at their defaults.

Question	Grok (xAI / SpaceXSI)	Claude (Anthropic)
Go-live date	7 Nov 2026, seven models	3 Nov 2026, three models; 16 Nov 2026, four more
Opt-out	Yes, team admins, per model	No opt-out for customers
Scope	Developer API only	Across products, including API, apps and Claude Code
How to check	x-watermark header, true or not	Detection API in private preview
Known limit	Detection access by application only	Older models need default sampling settings
Stated reason	EU AI Act compliance	EU AI Act December deadline

For background, see our earlier comparison, AI text watermarking Claude Grok. This article is the launch-day update: Grok's mark is live today, and Claude's second wave has a firm date of 16 November.

What This Does Not Do

A watermark is not a lie detector, and it is not a guarantee. It does not prove who wrote a text, only that the text matches a statistical pattern. Short texts give a weak signal, and the detector itself can return "uncertain". Paraphrasing or heavily editing a watermarked text can weaken or remove the signal. The Grok docs say nothing about Grok Imagine images or video, or about the Grok apps or Grok on X — this change is text, through the API, only.

There is also a trust question no provider has answered. The detector belongs to the same company that embedded the mark, and access to it is restricted, so no reader can independently check a suspicious text today. Until detection opens up, the watermark is mostly useful to the provider itself.

Illustration created for INORYVA

What Developers Should Do Today

First, check the header: read the x-watermark header on your Grok API responses from today and confirm it reads true for the seven listed models. Second, decide about opting out with legal advice. Third, do not build a product that depends on detecting watermarks in other people's text — you cannot get detection access without an application, and the signal is statistical, not certain. Fourth, if you use older Claude models, check whether your code changes temperature, top_p or top_k from their defaults.

The same trust question sits behind the Cursor Claude Code extension malware incident and the Claude rogue behaviour report.

Key Takeaways

• Grok API text carries an invisible watermark from today, 7 November 2026, on seven models from grok-4.20 to grok-4.7, on by default, to comply with the EU AI Act.
• The only per-response check is the x-watermark header: true means watermarked. Detection is statistical, stronger on longer text, and access is by application only.
• Team admins can opt out per model in the SpaceXSI Console, within 5 minutes, but must confirm a legal exception applies.
• Claude's first wave went live 3 November (Mythos 5, Opus 4.7, Sonnet 4.6); the second wave lands 16 November (Opus 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5), with no customer opt-out and a default-sampling caveat on older models.
• The Grok change covers the developer API only — no date for the Grok app, website or X — and it does not cover Grok Imagine images or video.
• Developers should check the header today and not build products that depend on detecting watermarks in other people's text.

Conclusion

Today AI-generated text starts carrying a hidden mark by default on the Grok API, and in nine days Claude's older models join in. Grok gives team admins a switch and a header to check; Claude gives no switch at all, and a watermark that depends on default settings. Neither system lets an ordinary reader verify a suspicious text yet. But the deadline is real: 2 December 2026 is when the EU AI Act's marking rule fully bites.

Sources

• SpaceXSI (xAI) developer docs, "Output Watermarking", updated 31 Oct 2026 — https://docs.x.ai/developers/advanced-api-usage/watermarking
• The Decoder, "Anthropic's text watermark is easy to bypass on several older Claude models", updated 4 Nov 2026 — https://the-decoder.com/anthropics-text-watermark-is-easy-to-bypass-on-several-older-claude-models/
• Notebookcheck, "Claude now watermarks text from older models, with no opt-out", 7 Nov 2026 — https://www.notebookcheck.net/Claude-now-watermarks-text-from-older-models-with-no-opt-out.1444769.0.html
• Regulation (EU) 2024/1689 (EU AI Act), Article 50 and Article 111 — https://eur-lex.europa.eu