Grok vs Claude: Two Ways to Watermark AI Text

Neon cyber image in hot pink, cyan and black comparing Grok and Claude AI text watermarking.

 

Two of the biggest AI providers are now watermarking their text — differently. xAI switched on Grok API watermarking on 7 November 2026; Anthropic began watermarking Claude text on 3 November, more models on 16 Nov. Both share one driver: EU AI Act Article 50(2), requiring machine-readable, detectable marking. The deadline is 2 December 2026.

What Grok is doing

From 7 November 2026, text from seven Grok models via the API is invisibly watermarked by default: 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 — covering text, structured outputs and tool-call arguments on both the Responses and Chat Completions APIs.

No code change is needed: the response format is unchanged except for a new x-watermark header (true = watermarked). The watermark adds no characters or hidden text; xAI says it does not affect quality or speed, though untested.

Team admins can opt out per model in the SpaceXSI Console, under Team settings and "Model output watermarking" (SpaceXSI is the xAI unit's name). Opting out requires confirming a legal exception applies and accepting sole responsibility. The Grok app, website and Grok on X are not covered.

What Claude is doing

Anthropic's rollout comes in two waves. The first, on 3 November, covered Claude Mythos 5, Opus 4.7 and Sonnet 4.6; Opus 4.6 was planned for 3 November but delayed, with no reason given. The second, on 16 November, covers Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5; post-2 Aug 2026 models already shipped with watermarking.

Claude's method is a variant of Google DeepMind's SynthID-Text — an imperceptible pattern in word choice, adding no tokens or hidden characters and carrying no information about the user, organisation or conversation.

The Claude loophole

Claude customers cannot switch the watermark off, which sounds stricter than Grok. But on the four second-wave models, and on Opus 4.6 and Sonnet 4.6, the watermark applies only with default temperature, top_p and top_k; changing any one may remove it. Anthropic's advice: "If you rely on the watermark for your own compliance, keep these settings at their defaults." No such caveat was stated for Mythos 5 and Opus 4.7.

That is a real loophole: developers change these settings routinely, and any change may strip the watermark.

The comparison

Grok (xAI / SpaceXSI)Claude (Anthropic)
Start date7 Nov 20263 Nov 2026 (wave 1); 16 Nov 2026 (wave 2)
Modelsgrok-4.7, 4.6, 4.5, 4.3, 4.20-reasoning, 4.20-non-reasoning, 4.20-multi-agentWave 1: Mythos 5, Opus 4.7, Sonnet 4.6; wave 2: Opus 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5; post-2 Aug 2026 models already watermarked
On/off controlOn by default; admins can opt out per model in the SpaceXSI Console, confirming a legal exception; some models locked onCannot be switched off by customers
LoopholeNo such caveat; removal needs a deliberate per-model opt-outOn second-wave models (and Opus 4.6 / Sonnet 4.6), the watermark applies only with default temperature, top_p and top_k; changing any one may remove it
DetectionDetection API: watermarked / not / uncertain; access limited, not publicDetection API in private preview for eligible organisations; regular users cannot check text
ScopeAPI only, every endpoint; Grok app, website and Grok on X not coveredAPI models in the two waves; Bedrock may lag a few days

Which approach is stronger?

On paper, Claude's "cannot be switched off" sounds stronger. In practice, the sampling-settings loophole weakens it for the second-wave models: the watermark can vanish through an ordinary configuration change. Grok's approach is more honest — its watermark is on by default, and switching it off is a visible, deliberate act requiring a legal justification. Neither system is verifiable by ordinary users, since both detection APIs are restricted.

Key takeaways

  • Both providers are watermarking text ahead of the EU AI Act deadline of 2 Dec 2026.
  • Grok: on by default from 7 Nov for seven models; admins can opt out per model, confirming a legal exception.
  • Claude: first wave 3 Nov, second wave 16 Nov; cannot be switched off — but second-wave models watermark only at default temperature, top_p and top_k.
  • Claude uses a DeepMind SynthID-Text variant; Grok's adds no characters or hidden text.
  • Neither detection API is public; Grok is API-only, Claude's second wave may reach Bedrock late.

Conclusion

Grok and Claude have both met the EU deadline, differently. Grok chose a visible default with a deliberate, accountable opt-out; Claude chose no off switch, then made the watermark depend on settings developers change daily. As 2 December arrives, the question is whether the marking survives real use.

FAQ

Why are Grok and Claude watermarking their text now?

Both comply with EU AI Act Article 50(2), requiring machine-readable, detectable marking. Deadline: 2 December 2026; Grok started 7 November, Claude's first wave 3 November.

Can the Claude watermark be turned off?

Customers cannot switch it off. But on the four second-wave models (and Opus 4.6 / Sonnet 4.6), it applies only at default temperature, top_p and top_k — changing any one may remove it.

Can I check whether a piece of text is watermarked?

Not as a regular user: xAI's detection access is limited and not public, Anthropic's is in private preview for eligible organisations.

Sources

Related reading on Daily Tech & AI:

Researched with AI assistance and reviewed for accuracy.

Follow on WhatsApp

Follow Daily Tech & AI on WhatsApp: https://www.whatsapp.com/channel/0029Vb9XzdHDJ6H4TFNAl61r

Comments