What AI text watermarks mean for your documents

In August 2026, Anthropic announced that new Claude models would embed an invisible watermark in the text they generate, everywhere Claude is available, to meet its commitments under the EU’s Code of Practice on Transparency of AI-Generated Content (Anthropic, 2026). OpenAI has signed the same code and says that its goal is to extend provenance signals to text, though at the time of writing its verification tool covers images and audio only (OpenAI, n.d.). If you draft with a model, some of your text is now marked, and the share will grow.

What a watermark is, and isn’t

A text watermark isn’t a hidden character or a tag in the file. When a model chooses between words that would all do, a cryptographic key nudges the choice, and over a long enough passage the pattern of those choices becomes something a detector holding the key can recognize. Readers can’t see it, and nothing in it identifies you or your conversation with the model.

That makes it a different thing from an AI detector. A detector guesses from style, and guesses badly. A watermark is a deliberate signal placed in the text by the company that made the model, and it’s read by that company’s own tool. There’s no guesswork in the reading, but there are limits to what the reading can tell anyone.

What it proves

Less than you might expect. In Anthropic’s own words, a watermark “can only determine that Claude was likely involved with the content at some point. It cannot distinguish ‘Claude wrote this’ from ‘Claude heavily edited this’” (Anthropic, 2026). A marked passage might be a paragraph the model drafted from nothing, or a paragraph you wrote and asked it to tidy. The mark doesn’t distinguish between the two.

It’s also fragile in specific ways. Anthropic says that light editing probably won’t remove it, a complete rewrite will, and short passages or highly constrained ones, such as lists of facts or code, carry little or no signal (Ha, 2026). So a marked document tells a reader that a model was somewhere in the chain, and nothing about how much of the thinking was yours.

Who can check

Not the public, for now. Anthropic’s detection API is in private preview for what it calls eligible organizations, such as regulators and fact-checkers, with wider access planned (Anthropic, 2026). OpenAI’s Verify tool and API check images and audio for OpenAI’s own signals and aren’t designed to detect anything from other companies’ models (OpenAI, n.d.). There’s no single checker, and a tool created to detect one company’s mark cannot identify content created by another.

What this changes for you

Assume that anything drafted or edited by a model is marked. If a passage came out of a current model, treat it as carrying a signal, whether or not anyone can read it today. That’s a reason to be deliberate about what you keep, not a reason to panic.

Don’t ask anyone to strip it. We don’t take work aimed at lowering a detector score, and the same applies here. Rewriting a document word by word to shed a watermark makes the document worse, and if anyone ever asks how it was produced, you’ll have a much harder question to answer than you would have had by simply being upfront.

Keep your own record of how the document was created. The mark proves involvement, not authorship, so if a client or a regulator asks, rely on your drafting history and your version record to answer the question of authorship. That was true before watermarks, and it’s even more true now.

Decide your disclosure position upfront. Transparency obligations under the EU AI Act apply from August 2026, and many organizations already have their own rules. Whether you disclose model use in a proposal or a report is a decision for you and your compliance people, and it’s easier to make it before the document is written than once someone has started asking questions about its provenance.

What it means when you work with us

Every change we make to your document arrives as a tracked change from a named editor, with margin comments where a decision is yours. If you have a disclosure requirement, or a rule about which tools are allowed to be used on a specific document, tell us before we start and we’ll work within it. And if the draft came out of a model, we’ll edit it exactly as we would any other, because while the mark changes what a reader can learn about the document’s history, it doesn’t tell you anything about whether it’s any good.