Claude is adding invisible marks to their writing

Welcome back.

Anthropic is preparing to place invisible watermarks inside text produced by supported Claude models.

You will not see a label at the bottom of a document. There will be no symbol attached to a sentence. The writing should look and read as it normally does.

But underneath the words, Claude will create a pattern that software can look for later.

The change is part of Anthropic’s response to new transparency requirements under the European Union AI Act. New Claude models launched in the European Union on or after August 2 will support marking from launch. Anthropic says those marks will apply wherever the supported models are offered, including Claude, its API, Claude Code, Cowork, and cloud platforms.

Existing models are still being brought into the system.

This sounds like a technical product update. I think it points to a much more ordinary workplace question.

When AI helps create something, who is responsible for what it says?

A watermark may help us see that Claude was involved. It cannot tell us whether the document is thoughtful, accurate, useful, or worth someone else’s time.

For that, we still need a person willing to stand behind the work.

Why Anthropic is doing this

The reason begins with a simple problem. AI generated content is becoming easier to create and harder to recognize.

A convincing image, voice recording, article, or report can now be produced in minutes. The person receiving it may have no reliable way to know where it came from or whether a machine played a meaningful role in creating it.

European regulators want AI companies to make that involvement easier for software to identify. Article 50 of the European Union AI Act requires providers of generative AI systems to mark certain outputs in a machine readable form when technically feasible. Those transparency rules began applying on August 2, 2026.

Anthropic signed the related code of practice and is applying its marking system more broadly than Europe. When a supported Claude model is used, the mark is intended to travel with the text worldwide.

The goal is provenance. That simply means giving people more information about where something came from and what may have happened to it before they received it.

Anthropic plans to use two methods.

• Generated text will contain an invisible watermark.

• Supported image files will carry signed information showing that Claude processed the file and whether it may have been altered.

This is a reasonable response to a real problem. We need better signals as synthetic content fills inboxes, websites, internal documents, and social feeds.

But a signal is not the same as a verdict.

Anthropic says a detected mark only suggests that Claude processed the content. It does not prove that Claude originated every idea or sentence. The mark could appear after Claude proofreads, translates, summarizes, or reformats something a person wrote.

And if the text is short, heavily edited, translated again, or mixed with other writing, the mark may no longer be detectable.

Clay’s writing policy asks the better question

Last week, Varun Anand shared the AI writing policy Clay introduced across the company.

The policy focuses on what employees owe the people reading their work.

Its first principle is the most important: you must stand behind every idea and sentence you share. If your name is on the document, the document should represent your thinking, regardless of which tools helped you produce it.

The other principles build from there.

• Writing is thinking. Working through an argument helps you understand the subject. If AI allows you to skip that process completely, you may publish something you do not understand well enough to defend.

• Respect the reader’s time. Generating a long document from a short prompt and asking ten colleagues to read it transfers your effort to everyone else.

• Longer is not better. AI makes it cheap to add polished sentences that contribute very little. The writer still has to decide what deserves to remain.

I like this approach because it deals with the actual workplace problem. It allows people to use AI while making clear that the tool does not inherit responsibility for the finished document at work.

The problem is not that AI touched a document. AI may have fixed grammar, organized notes, challenged an argument, or helped someone explain a difficult idea more clearly.

The problem begins when nobody can explain or defend the finished work.

A Claude watermark may tell your company that AI was involved. Clay’s policy asks the more useful questions: Did the person think? Did they check the claims? Did they remove unnecessary material? Are they willing to own the conclusion?

Those questions apply whether a document was written with Claude, another model, or no AI at all.

A watermark does not replace an owner

Companies should pay attention to Anthropic’s change, especially if teams use Claude to produce customer communication, marketing, research, policies, or internal decisions.

But I would not respond by banning AI writing or requiring employees to treat every detected mark as misconduct.

I would create a simple writing standard.

Before a document is shared, the person responsible should be able to say:

• I understand and agree with every material claim.

• I checked the facts, sources, and numbers that matter.

• I can explain how the conclusion was reached.

• I edited the document for the reader, not merely for polish.

• I am accountable if the work is unclear or wrong.

For consequential work, the company should also keep a record of the sources used, the reviewer, and the approval decision. That is more useful than trying to calculate what percentage of the words came from a person.

Watermarking will still matter. It can give platforms, publishers, and companies another clue about how a piece of content was produced. It may also make disclosure easier as regulations and customer expectations develop.

But we should be careful about what the mark actually tells us.

It can tell us that Claude may have processed the text.

It cannot tell us whether anyone thought carefully before publishing it.

That remains a management question, not a detection problem.

Clay’s policy gets the order right. Use whatever tools improve the work. Then make a person responsible for the result.

As AI becomes part of ordinary writing, that standard will matter more than whether every sentence can be classified as human or machine-made.

The useful question is simpler: who stands behind this?

Haroon

P.S. I am shaping what AI Ready covers next. If you have two minutes, share what would make this newsletter more useful to you.