Claude Now Marks Everything It Writes for You

If your team drafts with Claude, the words it hands back are no longer anonymous.

As of August 2, Anthropic began embedding an imperceptible watermark directly into text that Claude generates, and attaching signed provenance metadata to certain files it produces. In Anthropic's own description, the text mark is woven into the words themselves: "You won't see it, and it doesn't change the meaning, quality, or readability of Claude's response." It survives copy and paste. There is no opt-out documented, and the coverage is broad, spanning the API, the Claude apps, Claude Code, and the versions offered through AWS, Google Cloud, and Microsoft Foundry.

For files, the mechanism is different. Supported formats such as .svg, .png, and .jpg get signed metadata following the C2PA provenance standard, which records that the file was processed by Claude and reveals whether it was altered afterward.

Most business owners we talk to reacted to this news one of two ways. Either it sounded like a non-event, or it sounded like a trap. It is neither, and the difference matters for how you handle work product going out the door this quarter.

What the mark actually proves

Read Anthropic's caveats carefully, because they are doing a lot of work.

A detected mark "provides a signal that content was processed by Claude, but is not fully conclusive." And in the other direction: "Lack of a detected mark doesn't mean the content wasn't AI-generated or processed." The signal also weakens when text is heavily edited, paraphrased, translated, or converted between formats, and short passages may not carry enough of it to read at all.

So the mark establishes that a Claude model touched something. It does not establish who wrote what, how much was machine-drafted, or whether a human exercised real judgment over the result. A document that a person outlined, that Claude expanded, and that three colleagues revised carries the same kind of signal as one generated in a single pass and shipped untouched.

That gap between what the mark proves and what people will assume it proves is where the risk lives. Not in the technology. In the inference someone draws from it later.

Why this arrived now

This was not a spontaneous act of corporate conscience. Anthropic joined OpenAI and Google in outlining compliance with transparency requirements under the European Union's AI Act, whose Article 50 obligations took effect on the same August 2 date.

Under those rules, as summarized by the law firm Cooley, providers of generative systems must embed machine-readable markings in synthetic audio, images, video, and text, and provide a detection mechanism. Noncompliance can draw fines up to 15 million euros or 3% of worldwide annual turnover, whichever is higher. Providers of systems that existed before the deadline have until December 2, 2026 to finish the marking and detection work.

The useful takeaway is directional. Provenance marking is becoming table stakes for the major model vendors, not a differentiator. Assume the other tools your staff use are on the same path, whether or not they have announced it.

Four things worth doing before this becomes someone else's question

Find out what you have already promised. Some client contracts, RFP responses, and vendor questionnaires contain language about AI use, and a fair number of those clauses were signed without much thought. Go read yours. If you have certified that a deliverable was produced without generative AI, and it was drafted in a marked tool, you want to discover that yourself rather than have a client raise it.

Decide what your disclosure practice is, and write it down. Not a philosophical position. One paragraph covering which categories of work may be AI-assisted, what gets human review before it ships, and what you tell a client if asked. Firms that can answer this in a sentence look competent. Firms that improvise look evasive, and evasiveness is what turns a small issue into a large one.

Do not build anything on removing the mark. A cottage industry of watermark removers appeared within days of the announcement. Set aside whether they work, and most cannot, because Anthropic has not released the detection tools that would let anyone outside the company verify a claim either way. The reputational position of having stripped a provenance signal from client work is considerably worse than the position of having used AI in the first place.

Ask your vendors the provenance question directly. When you buy or renew a tool that generates text, images, or code, ask what it marks, what standard it follows, and who can read the mark. This is a reasonable procurement question now. The answers will tell you something about the vendor's maturity beyond this one issue.

The honest read

We think this is a good development that will be misread for a while.

Content provenance infrastructure is genuinely useful. In a few years, being able to establish where a document came from will protect the people producing honest work far more than it exposes them. The awkward stretch is right now, while the marks exist but the detection tools do not, which means everyone can be told the mark is there and nobody outside the vendor can check.

During that stretch, the durable protection is not technical. It is having done the work properly and being able to say so. If a client asks whether AI touched their deliverable, the answer that holds up is a clear description of how your firm uses these tools and who reviewed the output, given without hesitation.

The mark is only a problem for people whose process would not survive the question.

If you are working out what your firm's AI disclosure practice should say, that is the kind of question we work through with people.