Anthropic says future versions of Claude will start putting a hidden mark inside the text they write. You will not see it. Your customer will not see it. But software with the right key can check a piece of writing and estimate how likely it is that Claude helped produce it.
Here is how Anthropic puts it: "Future Claude models will generate text that contains a watermark. This is a way of determining the likelihood that Claude was involved in writing the text."
If you run a small shop and you have been quietly using AI to draft emails, proposals, and posts, this changes the math on hiding that fact. Not in a scary way. In a practical way. Let me walk through what it is, where it shows up, and what a busy owner should actually do about it.
What is Anthropic's Claude text watermark?
It is an invisible, machine-readable mark built into the words Claude picks. Anthropic changes the source of randomness Claude uses when it chooses among equally valid next words, settling each choice with a key plus the words that came before. A reader cannot perceive the pattern. Anyone with the key can verify it.
That last part is the interesting bit. The mark lives in the text itself, not in hidden characters and not in added metadata. So it travels with the words.
Anthropic's support page says it plainly: "Because the watermark is part of the text, it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing." Copy a Claude paragraph, paste it into a Word doc, and the signal can come along for the ride.
Where does the Claude watermark show up?
The watermark is applied at the model level. That means it is present no matter which Claude product the text came from. Anthropic lists the surfaces it covers: Claude, the Claude Platform (its API), Claude Code, Claude Cowork, and Claude Tag, including deployments run through AWS, Google Cloud, and Microsoft Foundry.
New Claude models launched in the EU on or after August 2, 2026 include this marking support at launch. Support for older, existing models is in progress and rolling out over the coming months during the AI Act transition period.
An admin who runs a two-person shop, a solo founder writing her own proposals, a marketing lead pasting a Claude draft into a client email: all of them now produce text that may carry a machine-detectable mark, no matter which Claude surface they used. Same signal, different door.
Images work differently. For three supported file types, .svg, .png, and .jpg, Claude does not use the text watermark. Instead it attaches signed provenance metadata that follows the Coalition for Content Provenance and Authenticity (C2PA) open standard.
Does the watermark identify you or your business?
No. Anthropic is direct about this: "There's nothing in the watermark, or its key, that would allow anyone to recover any information about the user." The company says it cannot be traced to a specific person, organization, or chat.
So the mark answers one question and one question only: how likely is it that Claude helped write this? It does not say who typed the prompt. It does not name your company. It flags likely AI authorship, not identity.
There is a limit worth knowing on the other side too. A missing mark is not proof a human wrote something. Anthropic warns that "lack of a detected mark doesn't mean the content wasn't AI-generated or processed." Model age, editing, brevity, stripped metadata, or platform limits can all wipe the signal. So detection can suggest AI involvement. It cannot prove a clean human record.
Why is Anthropic doing this, and does it cost anything?
Anthropic ties the move to its commitment under the EU Code of Practice on Transparency of AI-Generated Content, which it signed in July 2026 alongside roughly 190 total signatories. August 2, 2026 is the date from which the relevant EU requirement to mark AI-generated content applies. Anthropic's own materials point to global, model-level application, so this is not a Europe-only setting buried in a menu.
For you, the running cost is close to nothing. Anthropic says: "Watermarking has a negligible impact on the speed of models, and because it produces no extra tokens, the model is the same price to serve and use." Your Claude bill does not go up because your text now carries a mark.
One more thing is coming. Anthropic says it will soon offer a watermark detection API, and that it is still working out the details of how that will function. Once tools can check for the mark at scale, "did AI write this?" becomes a question software can start to answer.
How should you handle AI-drafted work now?
Two habits carry you through this, and both are within reach this week.
The first is disclosure. If a client or a platform expects to know when AI helped, tell them. That choice is far safer than betting that a paste will scrub every trace, especially once a detection API exists and the same signal appears across every Claude surface.
The second is editing. Read what Claude drafts. Push back on the weak parts. Rework it in your own words until it says what you mean and sounds like you. In the 7 Levels of AI Proficiency, that habit describes Level 3: Lieutenant, the Critical Thinker, the person who treats an AI draft as raw material and puts real judgment on top of it. You grow into that level through reps, not through hiding the tool.
Owning the output beats disguising it. A watermark rewards the owner and trips up the one trying to pass off an untouched draft as their own.
Your next step is small. Pick one place where you already lean on Claude, maybe client emails or proposals, and decide how you want to disclose and edit that work going forward. Write the rule down. Then you are ready for the day a detection tool asks the question for you.
Related reading: Level 3: The Lieutenant (Critical Thinker).
Sources
- Claude text watermark announcement (Anthropic)
- Text watermarking and content provenance (Claude support)
Frequently Asked Questions
Does editing remove the Claude watermark?
It might. Anthropic says the mark "may persist through some editing," but it also lists editing, brevity, and metadata stripping as reasons a signal can disappear. Primary sources do not quantify how much editing defeats it. Treat detection as a likelihood estimate, not a guarantee either way.
Can the watermark trace text back to me or my company?
No. Anthropic states nothing in the watermark or its key can recover information about the user, and it cannot be tied to a specific person, organization, or chat. It flags likely Claude authorship, not who you are.
Are AI-generated images watermarked the same way?
Not with the text watermark. For .svg, .png, and .jpg files, Claude attaches signed provenance metadata following the C2PA open standard instead.
Bring us the work that keeps coming back to you.
We will trace the workflow, identify where a system can create capacity, and decide what should remain with your people.