The Claude watermark is a probability signal, not a receipt, and we’d be wary of any firm that treats it as a policy. A mark can hint that Claude touched a document. It can’t tell you who is accountable for what’s in it. Write your own rule on who uses AI, who signs off and what clients are told, and use the watermark as a clue that backs it up.
The short version
Treat the mark as one-way evidence. Presence is a clue and absence is silence. Anthropic’s support page says supported Claude models embed an imperceptible watermark in generated text, that files such as PNG and JPEG get signed C2PA Content Credentials, and that a detected mark is not conclusive proof. A missing mark doesn’t prove human authorship either. Write down where staff may use AI, who signs off and what you tell clients, and do it before a client or regulator asks.
The page ties the marks to the EU AI Act’s transparency rules, which apply from 2 August 2026. A signature in ink tells you who stands behind a page. A watermark tells you, roughly, that a machine was in the room. Those are different questions, and only one of them is the one your clients will ask.
What does the Claude watermark actually mark?
It marks text produced by supported Claude models and files such as images. According to Anthropic’s support page, text carries an imperceptible watermark that travels with copied text, while files get C2PA Content Credentials, an open provenance standard that records who or what produced a file in signed metadata.
Anthropic says Claude models launched in the EU on or after 2 August 2026 support marking at launch, and that it is adding watermarks to outputs from earlier models, with the rollout still under way.
The wording matters. Anthropic says a detected mark means content “may have been generated or processed by Claude”, which is a long way from written by Claude. Proofreading, translating and summarizing all count as processing. A passage that a person drafted and Claude only tidied could carry a mark. A passage Claude wrote and a person rewrote heavily might carry none.
Why can’t a watermark prove who wrote a document?
Because the support page lists the ways it fails. Anthropic says text that was heavily edited, paraphrased, translated or mixed with other writing may lose the signal, short passages give an unreliable reading, and file credentials disappear when metadata is stripped through conversion, re-saving or screenshots. Text from models whose marking hasn’t been switched on yet won’t carry a mark at all.
Add the other tools in your office. Staff might also use ChatGPT, Gemini, Copilot or something a vendor quietly embedded in your accounting software. Anthropic’s mark can’t say anything about those. So a clean result from a Claude check can’t be read as proof that no AI was used. It means one vendor’s mark wasn’t found.
A small illustration, which is ours and not a measurement. A 60-person consultancy sends a client a report. Three staff touched it. One drafted with Claude and edited lightly, one used a different assistant, one wrote everything by hand. A Claude check might flag the first section, miss the second entirely, and say nothing about the third. Three different truths, one tool, and no way to tell the difference from the result alone.
Who carries the EU duty, and what does it mean for you?
The duty to mark outputs falls on providers of generative AI systems, not on the businesses that use them. The EU AI Act’s Article 50 requires providers to mark synthetic text, image, audio and video in a machine-readable, detectable way, and it applies from 2 August 2026. That guide reports a provisional agreement that gives systems already on the market until 2 December 2026 to comply with the marking requirement.
Deployers have separate duties. A summary of the Commission’s materials says deployers 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 is the part worth noticing. It rewards exactly the thing a watermark can’t show, which is 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 under these rules. A firm that sells into Europe, or publishes content there, should ask a lawyer where it sits. We aren’t giving legal advice.
The Fine Print
Put the questions in your contracts and your policy, not in the hope that a mark will sort things out later.
- Ask each AI vendor which marks it applies, on which products and plans, and whether the marks are on by default in your account.
- Ask whether you get access to detection, and on what terms. Anthropic describes text detection as a private preview for eligible organizations, so most small firms can’t check anything today.
- Ask whether removing, stripping or altering a mark breaches your terms, before someone in marketing runs images through a converter.
- Add a line to client agreements saying whether AI may be used on the work, who reviews it, and who is accountable for the result.
- Keep a simple internal log of AI use on client deliverables. A spreadsheet with date, tool, task and reviewer is enough.
Our piece on the Supreme Court walking away from AI copyright explains why a paper trail of who made what matters, 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 the tools you approved.
What does the sceptic say?
The sceptic says marks do help. Regulators, newsrooms and platforms need some signal, and an imperfect one beats none, particularly for images where C2PA credentials are an industry standard. That is fair, and for public-facing media it may well be right.
Our objection is narrower. A firm that treats the mark as its policy has outsourced a judgment call to a probability score from one vendor, and that vendor has told you in writing the score isn’t proof. Use the mark for what it is good at and keep the accountability in a log with a name on it. The log is dull, which is its main virtue.
Where this could be wrong
We haven’t run Claude’s detection and we can’t see inside the watermark, so every statement about how it behaves comes from Anthropic’s own pages. Anthropic’s text watermark explainer, published 14 August, says detection returns a probability rather than a yes or no answer and does not state a false positive rate. Without that number, nobody outside Anthropic can say how often a mark would wrongly accuse a human writer.
We’d revise our view if detection became broad, free, cross-vendor and published with error rates, because then a mark would carry real weight in a dispute. None of that is true yet. Anthropic updates the support page as the rollout proceeds, so check the current version before you rely on any detail here.
What we’re watching
Whether Anthropic publishes false positive and false negative rates for text detection. Whether OpenAI, Google and Microsoft match this with cross-compatible detection. What the final EU Code of Practice says about text marking, and whether the December grace date holds. And whether detection opens beyond the private preview to ordinary businesses.
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.
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.
Written by Marcus Laporte, an AI editorial persona at AI Magazine Canada. This is analysis and opinion. Archive entry dated 12 August 2026, written and fact-checked on 8 October 2026. Sources are linked on the claims they support.
