Claude and Grok Now Watermark AI Text: What It Means for Your Business Content

I use AI for a lot of my work, and I've never hidden that. So when two of the model makers I use said they're marking the text their models write, my first question was practical: does this change anything for the small businesses I build for? Mostly no. But a few details matter, especially if you publish for people in the EU or build AI into your own product.

Claude. Anthropic's explainer says it adds an invisible watermark to text from Claude models to comply with the EU AI Act, using a version of Google DeepMind's SynthID-Text. Its help article says Anthropic has signed the EU's Article 50(2) Code of Practice on marking AI-generated content. The watermark applies across the API, the Claude apps, Claude Code, Claude Cowork and Claude Tag, worldwide, and to supported models used through AWS, Google Cloud or Microsoft Foundry.

As of November 1, Anthropic's table lists text watermarks on Fable 5.1, Fable 5, Mythos 5.1, Opus 5.5, Opus 5, Opus 4.8, Sonnet 5.5, Sonnet 5 and Haiku 5.5. It says it's adding watermarks to models released before August 2, 2026, without giving a public date.

Grok. SpaceXSI's Output Watermarking docs, last updated October 31, say that starting November 7, 2026, text from these API models is 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

It's on every SpaceXSI API endpoint, including regional ones, and every text generation API, including the Responses API and Chat Completions. It covers normal replies, structured outputs and tool call arguments. SpaceXSI says there's nothing to change in your code and no effect on quality or speed. The page is about the API. It doesn't say anything about the Grok apps.

Neither company adds characters, markup or hidden text. Anthropic says there are "no hidden characters." SpaceXSI says the watermark "slightly shapes which tokens the model picks." In plain terms, when the model has several equally good next words, it leans toward some of them in a pattern only the vendor's detector knows to look for. One sentence doesn't show much. A few paragraphs add up.

That has three practical effects:

• You can't see it or strip it with a cleanup tool. Copy and paste keeps it. There's nothing to delete.
• Editing weakens it, but light editing probably isn't enough. Anthropic says "light editing probably won't remove the watermark completely; a complete rewrite where every word is replaced will." SpaceXSI hasn't published anything on how its mark survives edits.
• Detection is a probability, not proof. SpaceXSI says its detector answers watermarked, not watermarked or uncertain, and that "longer text gives a stronger signal."

Anthropic also says the watermark carries nothing that identifies you, your company or your conversation, and that it doesn't change who owns the output or your rights under its terms.

Files are handled separately. When Claude creates a file like a PNG or JPEG, Anthropic attaches signed Content Credentials, an open standard called C2PA, and offers a free Claude Content Checker to read them.

Not your competitors, and not your clients. At least not yet.

• Anthropic: detection is in private preview for organizations named under EU law, like regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and EU civil society groups, plus companies that need it for their own compliance. Anthropic says it plans to widen access.
• SpaceXSI: "Access to watermark detection is granted by application," through a form linked in its docs.

So the realistic risk isn't a client running your homepage through a checker tomorrow. It's that a lightly edited AI draft can likely be traced back to the model later, by people with access.

Article 50 of the EU AI Act is about transparency. The European Commission's FAQ says the marking and labelling rules apply from August 2, 2026, with a transition until December 2, 2026 for AI systems that were already on the market. The watermarks are the providers' side of that: the companies making the models have to mark output in a machine-readable way.

There's a second side for the people publishing. The Commission's code of practice also covers labelling AI-generated text that's published to inform the public on matters of public interest, and it leaves room for human review and editorial responsibility. Anthropic tells anyone building Claude into a product to independently assess what Article 50 requires of them.

I build software. I'm not a lawyer. If you publish for an EU audience or sell an AI feature there, get real legal advice. Don't take it from a blog post, mine included.

Yes, if you're a team admin and the model isn't locked on. Team admins can opt out per model in the SpaceXSI Console, under Team settings, "Model output watermarking." The setting covers every API key on the team and applies to new requests within 5 minutes. Opting out "asks you to confirm that an exception under applicable law applies to your use, and that you are solely responsible for any marking, disclosure, or other legal requirement." Some models may be locked on. Every change lands in the team's audit log.

So it's a legal decision as much as a settings one. If you're weighing it, talk it through with whoever handles compliance for you.

Anthropic doesn't describe an opt-out on its pages.

SpaceXSI adds one response header. x-watermark: true means that response was watermarked. Any other value, or no header, means it wasn't. From SpaceXSI's docs:

curl -i https://api.x.ai/v1/responses \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "grok-4.7", "input": "Write a haiku about the ocean."}'

Look for x-watermark in the headers. In Python with the OpenAI SDK, the docs use the raw response:

import os
from openai import OpenAI

client = OpenAI(api_key=os.getenv("XAI_API_KEY"), base_url="https://api.x.ai/v1")
raw = client.responses.with_raw_response.create(model="grok-4.7", input="Write a haiku about the ocean.")
print(raw.headers.get("x-watermark"))
print(raw.parse().output_text)

If you log responses for compliance, that header is worth saving alongside the output. That part is my suggestion, not SpaceXSI's.

• Website copy, blog posts, emails. If an AI wrote the first draft, assume the mark can still be there after your edits. That's fine. Google says it aims to reward "original, high-quality content" however it's produced, and that using AI mainly to manipulate rankings is spam. The watermark doesn't change that. Quality and honesty still do the work.
• Client work. Tell clients when AI helped. Edit it into their voice, because their name is on it. Don't promise anyone "100% human" copy if it isn't.
• Proofreading your own writing. Anthropic says when Claude only proofreads, there's "very little (if anything)" for the watermark to attach to.
• Code. Anthropic says code gets "generally less watermarking," because exact code leaves the model few free word choices, but "comments within code" can carry it. I'd assume the same for READMEs, docs and commit messages, since they're ordinary prose. Neither company has said that, and SpaceXSI hasn't said how much signal code carries. Its watermark does cover tool call arguments, which is how coding agents usually hand over the files they write.
• Your own AI features. If your app calls Claude or Grok and shows the text to users, that text can carry the mark. Check the x-watermark header on Grok responses, and read Article 50 with a lawyer if you have EU users.

Not much changes for me. I use Grok 4.7 almost all day for coding, and I've written about my AI coding setup. I'm going to keep telling clients when AI helped. I'll update this post if either company changes its model list or opens up detection.

If you want help setting up an AI workflow for your business that you're comfortable being upfront about, let's talk.

Does Claude watermark its text?
Yes. Anthropic says text from supported Claude models carries an invisible watermark based on SynthID-Text. As of November 1, 2026, that's Fable 5.1, Fable 5, Mythos 5.1, Opus 5.5, Opus 5, Opus 4.8, Sonnet 5.5, Sonnet 5 and Haiku 5.5, with older models being added.

When does Grok start watermarking text?
SpaceXSI says Grok API text from grok-4.7, grok-4.6, grok-4.5, grok-4.3 and the three grok-4.20 models is watermarked by default starting November 7, 2026.

Can editing remove an AI text watermark?
Anthropic says light editing probably won't remove Claude's watermark completely, while replacing every word will. SpaceXSI hasn't said how its watermark holds up to edits.

Who can detect AI text watermarks?
Only approved groups for now. Anthropic's detection is in private preview for organizations named under EU law and companies with compliance needs. SpaceXSI grants detection access by application.

Does the EU AI Act require AI text watermarks?
Article 50 requires providers of generative AI to mark output in a machine-readable way. The European Commission says those rules apply from August 2, 2026, with a transition to December 2, 2026 for systems already on the market. Ask a lawyer what applies to you.

Will watermarked AI content hurt my Google rankings?
Google says it rewards original, high-quality content however it's produced, and treats AI used mainly to manipulate rankings as spam. It hasn't said anything about these watermarks.

• Anthropic, "How Claude's text watermark works" (August 14, 2026, updated September 1)
• Claude Help Center, "How Claude marks AI-generated content"
• Claude Content Checker
• SpaceXSI Docs, "Output Watermarking" (last updated October 31, 2026)
• European Commission, FAQ on the Code of Practice on Transparency of AI-Generated Content
• Google Search Central Blog, "Google Search's guidance about AI-generated content" (February 8, 2023)