AI Text Watermarking Claude Grok: One Reported Date, One Documented Date

Anthropic and xAI are both preparing to watermark AI-generated text, but the proof behind each plan differs. Grok's 7 November 2026 start date is written in xAI's own public documentation. Claude's 3 November 2026 start date for four older models comes only from a customer email reported second-hand; Anthropic's public support table, checked on 2 November 2026, does not show it yet. Neither rollout had started when this article was written. That difference is the story: a mark readers cannot check is only as useful as the record behind it.

Both call the work EU AI Act compliance and the mark invisible in the text. Anthropic calls it imperceptible, woven into the text; xAI says it shapes token picks, adding no characters, markup, or hidden text. Checking a passage for the mark is restricted in both cases.

Claude: the public table vs the reported email

Anthropic's support page, "How Claude marks AI-generated content" (updated 28 October 2026; fetched 2 November 2026), has three columns: text watermark on Anthropic's surfaces, text watermark via cloud partners (AWS, Google Cloud, Microsoft Foundry), and Content Credentials (C2PA) in files. Nine models show all three marks: Claude Fable 5.1, Claude Fable 5, Claude Mythos 5.1, Claude Opus 5.5, Claude Opus 5, Claude Opus 4.8, Claude Sonnet 5.5, Claude Sonnet 5, and Claude Haiku 5.5.

Seven show Content Credentials only, with no text watermark in either text column: Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.6, Claude Sonnet 4.5, and Claude Haiku 4.5. The page states: "Consistent with our commitments under the Code, Anthropic is adding watermarks to outputs from models released before August 2, 2026." It gives no rollout date and no mention of temperature, top_p, or top_k.

ai-text-watermarking-1.jpg

Illustration created for INORYVA — abstract text lines with an invisible mark; no readable sentences, no vendor logo.

Applied at the model level, it travels with copy and paste. Detection "is currently in private preview, available to eligible organizations as required under EU law." The free Claude Content Checker covers C2PA files only, not text-watermark detection.

The four models and 3 November come from a different kind of source. The Decoder reported on 31 October 2026 that, in an email to API customers, Anthropic said it will watermark Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6, and that the Bedrock rollout may take a few extra days. The same reporting says customers cannot turn it off, calls it a SynthID-Text variant, and cites the EU AI Act December deadline. None of that is on the public page.

The date comes from secondary reporting. FourWeekMBA quotes an email headed "Text watermarking is coming to more Claude models": "On November 3, 2026, Anthropic will begin applying its text watermark to outputs from these four models" and "This is the same watermark our newer models, including Claude Opus 5.5, already carry." The sourcing note says the date and names come from the email as a screenshot posted on X on 1 November 2026; Anthropic has not published it.

According to The Decoder, the reported email also conditions the mark on sampling settings for only two models: on Claude Opus 4.6 and Claude Sonnet 4.6 it kicks in only when temperature, top_p, and top_k stay at their defaults, and a minor tweak to any one dodges it. The quoted warning: "If you rely on the watermark for your own compliance, keep these settings at their defaults." The email does not mention the caveat for Mythos 5 or Opus 4.7, which The Decoder reads as the mark applying regardless of settings — its inference, not a confirmation. The Claude Duet article covers a separate Anthropic product; the Claude Haiku 5.5 price article covers Haiku 5.5 pricing (already marked on the public table).

Grok: the date is in the public docs

xAI's "Output Watermarking" docs page (updated 31 October 2026; fetched 2 November 2026) says SpaceXSI watermarks text model output for EU AI Act compliance, enabled on every SpaceXSI API endpoint. The page says nothing about grok.com, the Grok apps, Grok on X, or Grok Bot.

The start date needs no second-hand source: "Starting November 7, 2026, responses from these models include a watermark. It's on by default, and you can opt out per model." The seven models are grok-4.7, grok-4.6, grok-4.5, grok-4.3, grok-4.20-reasoning, grok-4.20-non-reasoning, and grok-4.20-multi-agent. Secondary coverage citing an email screenshot on X reports the same details.

ai-text-watermarking-2.jpg

Illustration created for INORYVA — abstract two-lane comparison visual; no vendor logos, no screenshots, no readable text.

The watermark covers outputs from every text generation API, including the Responses API and the Chat Completions API — text responses, structured outputs, and tool call arguments. Developers can read the x-watermark header: x-watermark: true means that request was watermarked; any other value, or no header, means it was not. The header checks one API response, not arbitrary text.

The docs say watermarking "slightly shapes which tokens the model picks, so it adds no characters, markup, or hidden text," and that "Watermarking doesn't change response quality or speed." That is a vendor statement; no independent testing was retrieved. No request parameter exists; the response format is unchanged apart from the header.

The opt-out is per model and sits with team admins, who can turn watermarking off for some or all models in the SpaceXSI Console; the setting applies to every API key on the team. Only admins see the "Model output watermarking" row; changes apply within five minutes; every opt-in and opt-out is recorded in the team audit log. Opting out asks the admin to confirm a legal exception applies and the team bears sole responsibility for marking or disclosure. Some models may be locked on. Detection is statistical — watermarked, not watermarked, or uncertain, stronger for longer text — and "access to watermark detection is granted by application."

Side by side: two kinds of evidence

Every cell comes from the sources above. Claude mixes a public table with a reported email; Grok rests on one docs page.

Claude (Anthropic)Grok (xAI)
Documented today9 newer models: all three marks. 7 older: C2PA only.Docs: 7 models from 7 November 2026.
Start date3 November 2026, reported only, email seen on X.7 November 2026, in the public docs.
Evidence levelPublic table plus reported email; unchanged 2 November.Public docs, updated 31 October 2026.
Can it be turned off?Reported: customers cannot; Opus/Sonnet 4.6 at defaults only.Per-model team opt-out; some models locked on.
Detection accessPrivate preview; checker covers files only.By application; header checks one response only.

Grok's terms can be verified before the start; they are published. Claude's expansion may happen exactly as reported, but on 2 November it could not be verified from Anthropic's public page.

What neither watermark gives readers

Neither vendor gives the public a way to test a passage. Claude's detection is in private preview; Grok's is by application. The only public checks retrieved are Claude's Content Checker (files only) and Grok's per-response header. A reader handed a screenshot or pasted paragraph has no detector from either vendor.

ai-text-watermarking-3.jpg

Illustration created for INORYVA — magnifier over blurred, unreadable text; not a detector screenshot.

Neither mark proves authorship on the evidence retrieved. Anthropic's "may persist through some editing" covers copy, paste, and some editing — not paraphrase — and no independent survival testing was retrieved for either vendor. Searches for discussion tied to these announcements found nothing solid; no community comment is quoted. For a different vendor's governance approach, see the OpenAI Model Spec.

Key Takeaways

  • Grok's 7 November 2026 date is in xAI's public docs; Claude's 3 November 2026 date is email-reported only, and Anthropic's table did not show it on 2 November.
  • Claude's table: 9 newer models with all three marks, 7 older with C2PA only. The reported expansion names Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6.
  • As reported, Opus 4.6 and Sonnet 4.6 watermark only at default temperature, top_p, and top_k — absent from the public page, and its absence for the other two models is an inference, not a confirmation.
  • Grok's docs name 7 models, on by default from 7 November, a per-model team opt-out, and an x-watermark header for one API response.
  • Detection is restricted for both — Claude private preview, Grok by application — with no public text detector in the sources retrieved.
  • Neither mark is shown to prove authorship or survive paraphrase, and xAI's quality/speed neutrality is a vendor statement without independent testing retrieved.

Conclusion

Written 2 November 2026, this is a snapshot. Grok's terms can be checked in a public docs page today. Claude's expansion can only be checked against a reported email and an unchanged table. Until the table changes or Anthropic publishes the email: one date documented, one date reported, no public detector for either watermark.

Sources

Comments

Post a Comment