Anthropic now puts an invisible watermark in text that Claude writes. The company explained how it works in an August 14, 2026 post, a few days after TechCrunch first reported the plan, and updated it on September 1 with details on who can use the detector. The watermark exists to meet EU transparency rules, but Anthropic applies it everywhere, so it covers your business whether or not you ever sell in Europe.
If your blog posts or service pages pass through Claude, some of that text now carries a mark a machine can find. That doesn’t put the content in trouble. It does mean “nobody can tell” is no longer a safe assumption, so it helps to be clear about what the mark proves and what it doesn’t.
How Claude’s Text Watermark Works
Anthropic, along with other AI providers, signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026. TechCrunch reports that Google, Meta, Microsoft, and OpenAI have also committed to the code. Here’s what Anthropic’s post and its help center page say, in plain terms:
- Where it applies: The help center lists the Claude apps, the Claude Platform API, Claude Code, Claude Cowork, and Claude Tag, and says supported models accessed through AWS, Google Cloud, or Microsoft Foundry carry the watermark too. Anthropic applies it globally because it doesn’t yet have a “durable way to scope it by region.”
- Which models: Claude models launched in the EU on or after August 2, 2026 support the marking at launch. Older models are in a transition period, and Anthropic says watermarking is being added to them over the coming months.
- How text is marked: When Claude has a choice between similar words (Anthropic’s example is describing a cold day as “overcast” or “grey”), a secret key nudges which one it picks. Across enough text, those choices form a pattern a detector with the key can see. Anthropic calls it “a version of the SynthID-Text approach” that Google DeepMind published in 2024.
- How files are marked: Supported files such as PNG, JPG, and SVG images get a C2PA content credential instead, which Anthropic describes as “a small, cryptographically signed note in the file’s metadata.” The file itself isn’t altered.
- Effect on the writing: Anthropic says, “To a reader, a watermarked response is indistinguishable from an unwatermarked one.”
- Privacy: According to Anthropic, the mark “doesn’t identify anything to do with individual users.”
Who Can Check for It
Not you, and not the general public. The detection API is in private preview. The help center lists eligible groups as regulators, law enforcement, media, fact-checkers, independent researchers, and EU civil society groups, plus enterprises with their own compliance obligations. Anthropic says it plans to widen access and has an access request form. It hasn’t given a date for public access.
What the Watermark Can’t Tell Anyone
Anthropic lists the limits openly, and for a business they matter more than the feature itself:
- Detection “doesn’t work well on small samples.” A meta description or a two-line ad doesn’t give it much to go on.
- The watermark is sparser on factual passages, and it’s generally not applied where exact output is required, so code carries less of it.
- When Claude only lightly corrects your own writing, “there’s very little (if anything) for the watermark to attach to.”
- “Light editing probably won’t remove the watermark completely.” A complete rewrite will. The help center adds that heavy editing, paraphrasing, translation, or mixing the text into other writing can make it undetectable.
A positive result means Claude was likely involved somewhere. It doesn’t say how much, and it says nothing about whether the content is any good. The help center also points out that people use Claude to proofread, translate, and summarize, so a mark can show up on material Claude didn’t originally write.
Detection Is More Reliable, but the Risk Is the Same
AI detection used to mean classifier tools that guessed whether writing “sounded like AI,” and they were wrong often enough that nobody could rely on them. A watermark is a signal placed in the text on purpose, which makes a positive result much harder to wave away. What it can be used for is narrower than a lot of the coverage implies.
A Watermark Is Not a Ranking Penalty
Google’s position has been consistent since February 2023. Its guidance on AI-generated content says “Appropriate use of AI or automation is not against our guidelines.” What Google treats as spam is using automation “with the primary purpose of manipulating ranking in search results.” Google hasn’t said it would use this watermark for ranking, and nothing in Anthropic’s announcement involves search engines.
So a helpful, accurate page that Claude helped draft is in the same position it was in July. A thin page churned out to chase keywords is also in the same position it was in July: exposed. The watermark didn’t create that risk. Our SEO strategy work has treated unreviewed AI output as a liability for a long time, and this doesn’t change that.
The People Who Can Check Care About Trust
Look at the access list again: regulators, journalists, fact-checkers, researchers. Those are the people who might ask whether your published claims hold up. If a reporter checks a statement in your press release, or a regulator reviews marketing for a licensed profession, “Claude was involved” may now be something they can establish. That’s only a problem if the text says something you can’t stand behind.
This matters most for businesses that already answer to disclosure rules or professional standards, such as law firms and medical practices. It’s also worth a thought if your contracts promise clients “original, human-written” work.
Your Writers and Vendors May Already Use Claude
Plenty of businesses don’t know which AI tools touch their content. Freelance writers and agencies often use Claude or similar tools without mentioning it. TechCrunch reported that some users were upset the watermark could expose their use of Claude at work or school. If you pay for content, it’s reasonable to know how it’s made.
How to Handle AI-Assisted Content Now
- Find out where AI touches your content. Ask your writers and any agency which AI tools they use and on which pieces. You aren’t hunting for violations. You need an accurate picture, especially for anything that makes claims about your credentials or pricing.
- Require a real human review. Every AI-assisted page should be checked by someone who knows the subject: every fact and number verified, plus the things AI can’t know, like how you actually do the work. Whether a watermark survives the edit is beside the point, so skip the “AI humanizer” tools that promise to strip it out.
- Check your promises and obligations. If your contracts or site copy say your content is written entirely by people, make sure that’s true or change the wording. If you operate in the EU or work in a regulated field, ask your counsel whether any disclosure rules apply to what you publish.
If AI search is part of your plan, the same standard applies there. The pages that get picked up in Google AI Overviews are the ones with clear, accurate, specific answers, however they were drafted.
Proof of Involvement, Not Proof of a Problem
Claude’s text watermark can show that Claude was involved in a piece of text. It can’t show the text is bad, and it doesn’t hand Google a reason to demote it. The businesses with real exposure are the ones publishing unreviewed AI output and hoping nobody looks closely. The fix is the same as before: someone who knows your business reads and approves every AI-assisted page before it goes live. We’re watching how detector access expands and will update clients if search or ad platforms start using these signals.
If you’re not sure how AI fits into your content process, or whether your pages would hold up to a close look, schedule a free strategy call. We’ll walk through it with you.
