Claude Marks the Text It Writes. Decide What You Tell Clients.
From November 3 more Claude models carry an invisible watermark. It changes little in your day, but it settles one question every office should answer.
By John Martines · November 1, 2026 · 6 min
Anthropic is adding its invisible text watermark to four more Claude models on November 3. Here is what the mark is, what it is not, and the short disclosure rule a small business should write this week.
The office manager at a small dental practice asks Claude for a patient newsletter. She reads it, fixes two sentences, adds the new hygienist's name and sends it to 900 patients. The same afternoon a contractor's estimator uses Claude to tidy the scope section of a bid for a county job.
Neither of them sees anything different on the page. But from November 3, text like theirs carries a mark they can't see, and someone with the right tool could find it.
That is not a reason to worry. It is a good reason to decide, calmly and in writing, what your business says about the AI it uses.
A text watermark is a hidden pattern in the words themselves. When Claude picks between words that mean nearly the same thing, the watermark nudges which one it picks, following a secret key Anthropic holds. One word proves nothing. Across a few paragraphs, the pattern adds up to something a detector can measure.
Anthropic calls it an "imperceptible pattern in word choice." By its account it adds no hidden characters or extra tokens, doesn't change prices, and carries no information about you, your company or your conversations. It sits inside the model, so it is there whether you use Claude directly or through Amazon, Google Cloud or Microsoft, and a customer can't switch it off. Because it lives in the words, it travels when the text is copied and pasted, and it may survive light editing. A full rewrite, where nearly every word changes, removes it.
Two limits matter as much as the mark. Finding it doesn't prove Claude wrote the whole thing, since people often use Claude to edit or translate their own writing. And a missing mark doesn't prove a person wrote it.
How it got here:
- August 2, 2026: every Claude model launched from this date on carries the mark from day one, according to Anthropic's help center.
- August 11, 2026: Anthropic announced it would watermark the text its models generate, tied to the transparency rules in the European Union's AI Act, and applied it worldwide rather than only in Europe.
- October 2026: OpenAI announced its own text watermark, textGrain, starting with ChatGPT and Codex in the EU. For OpenAI's API it is opt-in and off by default. Google has used its SynthID text watermark in Gemini and released it as open source.
- Late October 2026: Anthropic emailed API customers that Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6 get the mark from November 3, with Amazon Bedrock a few days behind. On Opus 4.6 and Sonnet 4.6 it applies only at the default settings for how the model picks words.
- December 2, 2026: the date reported for the EU's machine-readable marking rule to apply to AI systems already on the market before August 2.
Who can check for the mark? For now, not your client with a free website. Anthropic's detection tool is in private preview, offered to groups like regulators, researchers and fact-checkers. That will likely widen over time, so plan as if anyone may check one day.
Most of the Claude text a small business sends out (a newsletter, a proposal, a job post, a reply to a review) is ordinary business writing. Nobody is going to run it through a detector, and nothing in the law requires a small office to label a customer email. The watermark doesn't change that.
It does change one thing: "nobody can tell" is no longer a safe plan. Three places to look.
Client and contract terms. Some clients, especially larger companies, government buyers and law firms, now put AI clauses in contracts: disclose it, get approval, or don't use it at all. Legal writers have already pointed out the risk for a firm whose client banned AI and later finds the mark in a document. If you signed one of those clauses, the watermark makes it checkable.
Schools, grants and applications. Scholarship essays, grant narratives, licensing exams and job applications are where detection is most likely to be used first. If your staff help with any of these, the rule needs to be clear.
Your reputation. Getting caught looks worse than saying so. A business that says "we use AI to draft and a person checks everything" rarely loses a customer over it.
1. Find your AI clauses. Ask whoever handles contracts to search your active client agreements and recent bid documents for "artificial intelligence," "AI" and "generative." Make a short list: which clients forbid it, which want notice, which want approval. Most small businesses will find none or one. Knowing that is the point.
2. Write a two-line disclosure rule. For example: "We use AI tools to help draft and edit. A person on our team reviews, corrects and approves everything we send." Then add the exceptions from step 1: "For clients A and B, no AI on their work." Put it in your staff AI rules. If you don't have those yet, the free [Staff AI Rules Kit](/tools/staff-rules) builds them in about ten minutes.
3. Don't buy a "humanizer." Tools that promise to scrub AI text so it can't be detected are selling the wrong fix. They make the writing worse, they don't change what your contract says, and using one turns an honest practice into something that looks like hiding. If a client must not get AI-assisted work, write it yourself.
4. Keep a person's name on what goes out. The watermark is a reminder of a rule that was already good: someone reads, corrects and owns every piece of AI-drafted text before it leaves the building. Facts, prices, names and promises get checked by a person every time.
5. If you build on Claude, tell your developer. If your website, chat assistant or app sends Claude-written text to customers through Anthropic's API, the mark now comes along. Anthropic told API customers that anyone relying on the watermark for their own compliance should keep the model settings at their defaults. Ask your developer whether your setup changes those settings.
- Do any of our active contracts or bids mention AI, and what do they ask for?
- Have we written down, in two lines, what we tell customers about AI?
- Does every AI-drafted piece have a named person who reviewed it?
- Is anyone on staff using AI for applications, exams or grant writing where detection may be used?
- If we use Claude through our website or software, does our developer know about the November 3 change?
Watermarks are becoming normal across the big AI providers, pushed along by European rules, and detection will get easier to reach. Expect more clients to ask "did you use AI on this?" in plain words, the way they now ask about insurance.
You don't need to change how you write this week. You need a short, honest answer to that question, and a list of the clients who want a different one. Write both down before someone asks.
---
Sources (checked November 1, 2026). How the mark works, what it carries, who can detect it, and its limits: Anthropic, "How Claude marks AI-generated content" (Claude Help Center, 2026). The original announcement and its EU tie: "Anthropic says it will watermark text generated by its AI models" (TechCrunch, Aug 11, 2026). Survival through copying and editing: "Anthropic's Claude will watermark AI-generated text. Here's how it works" (Global News, Aug 2026). The four models, the November 3 date, Bedrock timing, default settings on Opus 4.6 and Sonnet 4.6, and the law firm concern: "Anthropic's watermark rollout leaves a loophole on two older Claude models" (The Decoder, Oct 31, 2026) and Pasquale Pillitteri, "Claude watermark: November 3, Opus 4.7, Sonnet 4.6" (Oct 2026). OpenAI's textGrain and its EU rollout: "OpenAI will watermark ChatGPT text in the EU" (Mixed News, Oct 2026) and FourWeekMBA (Oct 2026). EU AI Act Article 50 marking date: Cloud Security Alliance, "EU AI Act Watermarking Grace Period Ends December 2026" (research note, Sept 21, 2026).