The Claude watermark won’t settle who wrote what, so write the rule yourself

The Claude watermark marks text and files, but it cannot prove who wrote what. Marcus Laporte sets out the policy and contract terms a firm needs now.
A hand holding a black fountain pen and writing on white paper.

Read the Claude watermark’s own wording first. Anthropic says a detected mark means content “may have been generated or processed by Claude.” Processing covers proofreading, translating and summarizing. So the mark is a clue that a machine touched the page. It records nothing about who is accountable for it, and a firm that builds its AI policy on the mark has handed that policy to a probability score.

The EU’s transparency rules for AI-generated content apply from August 2, 2026, and Anthropic now marks what Claude writes and makes. Write your own rule anyway: who may use AI, who signs off, what clients are told. Then let the mark back it up.

Key points

  • Text carries an imperceptible watermark; PNG and JPEG files get signed C2PA Content Credentials.
  • Anthropic lists the failure cases itself: heavy edits, paraphrase, translation, short passages, stripped metadata, and models not yet marked.
  • Article 50’s marking duty falls on providers. Deployers have separate disclosure duties, with an exemption where a named person reviewed the content.
  • A client agreement line and a four-column log answer the accountability question a mark can’t.

What the support page says, clause by clause

Scope. According to Anthropic’s support page, text from supported Claude models carries an imperceptible watermark that travels with copied text. Files such as PNG and JPEG get signed C2PA Content Credentials, an open provenance standard recording who or what produced a file in signed metadata. The page ties both to the EU AI Act’s transparency rules.

Advertisement

Timing. Claude models launched in the EU on or after August 2, 2026, support marking at launch. Anthropic is adding watermarks to earlier models’ outputs, and the rollout is still under way.

Limits. Heavily edited, paraphrased, translated or mixed text may lose the signal. Short passages give an unreliable reading. File credentials disappear when metadata is stripped through conversion, re-saving or screenshots. Text from models whose marking isn’t on yet carries no mark.

Detection. Anthropic’s text watermark explainer, published August 14, says detection returns a probability rather than a yes or no answer, and it states no false positive rate. Text detection is a private preview for eligible organizations, so most small firms can’t check anything today.

Why the mark can’t settle authorship

A passage a person drafted and Claude only tidied could carry a mark. A passage Claude wrote and a person rewrote heavily might carry none. And Anthropic’s mark says nothing about ChatGPT, Gemini, Copilot or the assistant built into your accounting software. A clean Claude check means one vendor’s mark wasn’t found.

Take a 60-person consultancy sending a client a report, our own illustration. One staffer drafted with Claude and edited lightly, one used a different assistant, one wrote by hand. A Claude check might flag the first section, miss the second, and say nothing about the third. Three different truths, one tool, no way to tell them apart.

The mark helps Anthropic show regulators it complied, and it helps platforms sort content. A client or a court will ask a different question: who is responsible for this sentence? A mark can contradict your answer or fail to support it. Only a person can give the answer, and only a record can prove it. Without a published error rate, nobody outside Anthropic can say how often a mark would wrongly accuse a human writer.

Who owes what under Article 50

The EU AI Act’s Article 50 requires providers of generative AI systems to mark synthetic text, image, audio and video in a machine-readable, detectable way, from August 2, 2026. The same guide reports a provisional agreement giving systems already on the market until December 2, 2026. Businesses that only use these systems carry no marking duty.

Deployers have separate duties. A summary of the Commission’s materials says they must disclose deepfakes and certain AI-generated text published to inform the public, with an exemption where a person has reviewed the content and accepts editorial responsibility. That exemption rewards exactly what a watermark can’t show: a named human who checked the work.

As we noted in our Washington frontier AI review piece, most AI rules reach you through your vendors. A Canadian firm with no EU customers may owe nothing here. A firm that sells or publishes into Europe should ask a lawyer where it sits. This is general information, and we give no legal advice.

Four questions for vendors, two documents for you

  1. Which marks does each AI vendor apply, on which products and plans?
  2. Are they on by default in your account?
  3. Do you get access to detection, and on what terms?
  4. Does removing or altering a mark breach your terms? Ask before marketing runs images through a converter.

Then the two documents. A line in client agreements saying whether AI may be used on the work, who reviews it and who is accountable. And a simple log of AI use on client deliverables: date, tool, task, reviewer. Dull on purpose.

Our piece on the Supreme Court walking away from AI copyright explains why that paper trail matters, and AI-generated work copyright now rests on it. Our look at the Munich ruling on AI disclaimers shows a court treating disclosure as no shield, and our look at shadow AI in Canadian workplaces explains why you can’t assume staff use only approved tools.

The case for marks is real. Regulators, newsrooms and platforms need some signal, and C2PA credentials are an industry standard for images. For public-facing media that may be enough. Our objection is narrow: Anthropic has said in writing that its score isn’t proof, so a policy built on it rests on a judgment the vendor disclaims. We’d revise that if detection became broad, free, cross-vendor and published with error rates. We haven’t run Claude’s detection ourselves, and Anthropic updates the support page as the rollout proceeds, so check the current version.

Name a reviewer for each deliverable by role, and start the log this month.

Frequently asked questions

Does the Claude watermark prove my staff used AI?

No. Anthropic says a detected mark suggests content may have been generated or processed by Claude and is not conclusive proof. It also says a missing mark does not prove human authorship.

Does my business have to mark AI content under the EU AI Act?

The marking duty in Article 50 falls on providers of generative AI systems. Deployers have separate disclosure duties for deepfakes and some published text. If you sell into the EU, ask a lawyer how those apply to you. OpenAI took a different route in Europe, and the ChatGPT watermark does not cover Canadian users yet.

What should a small business put in its AI policy about watermarks?

Name who may use AI on which work, who reviews it, and what you tell clients. Keep a simple log of AI use on deliverables, because a log records accountability and a watermark does not.

Archive entry dated August 12, 2026, written and fact-checked on October 8, 2026. Sources are linked on the claims they support.

Total
0
Shares
Prev
Bill C-36 Regulates Your AI Without Being an AI Law
Laptop on a round table showing a padlock and the word secured on screen.

Bill C-36 Regulates Your AI Without Being an AI Law

Bill C-36 replaces PIPEDA and reaches AI through privacy rules

Next
Canada Ranked 44th of 47 on AI Training and It Shows
A classroom filled with lots of desks and chairs.

Canada Ranked 44th of 47 on AI Training and It Shows

Canada ranks 44th of 47 on AI training and literacy while adoption triples

You May Also Like