When I first read that Claude’s output would be watermarked, I braced for hidden characters embedded in the text. Would pasting it into a blog make it identifiable? Would it affect search rankings? Would I need to strip it out afterwards?
No to all three, and the mechanism is not what I pictured. Anthropic’s watermark is a statistical pattern in which words get chosen during generation, checkable against a key by a detector that returns a likelihood. Nothing is added to the text. There are no hidden characters. It carries no identifying information.
For anyone publishing a blog, the practical answer is that almost nothing changes — and the one thing that is worth knowing is not about the watermark at all. It is that EU AI Act Article 50, in force since August 2, 2026, exempts published AI-generated text from its disclosure obligation where the content has undergone human review or editorial control and a person holds editorial responsibility for it. Which is to say: the thing that protects you was always the editing, not the marking.
Originally published in Japanese on August 26, 2026. This English version was written in September 2026 and corrects one point from the original.
What the watermark actually is
Anthropic announced on August 14, 2026 that future Claude models will generate text containing a watermark. This applies to ordinary text; it is not a visible mark like a logo overlaid on an image.
Per Anthropic’s explanation, the choice among candidate words is still random — what changes is the source of that randomness. Instead of an arbitrary random number generator picking the next word, the key and a few preceding words settle which word the model picks. The words Claude picks remain random, and the sequence can afterwards be checked for consistency with the choices Claude would have made using the key. The detector examines that pattern and assigns a probability that the text came from Claude.
Two statements from Anthropic are worth quoting the substance of exactly, because they dispose of the most common fears. Nothing is added to the text and there are no hidden characters. And the watermark carries no identifying information and cannot be traced to a specific person, organization, or chat.
So treating a watermark as a personal tracking mark is simply inaccurate.
It is a version of SynthID-Text — I was too cautious about this
In the Japanese original I was careful not to say the implementation was SynthID-Text, on the grounds that resembling a published technique is not the same as shipping it. That caution was misplaced, and I am correcting it here.
Anthropic states directly that Claude’s text watermark is a version of the SynthID-Text approach published by Google DeepMind in a 2024 Nature paper, and places it within a line of work going back to a 2022 proposal by Scott Aaronson.
The precise reading is “a version of the approach” — not “Google’s SynthID Text product has been installed into Claude,” and not “this merely resembles a paper.” Being vague in the safe direction is still being vague. When the official text is specific, match its specificity.
What detection can and cannot establish
A detection result is not a certificate of authorship.
Anthropic notes that detection does not work well on small samples, where there are fewer word choices and therefore less information to work with. Longer text yields more confidence. Translations produced by Claude do carry a watermark, because in that case every word is chosen by Claude. Conversely, heavily rewritten text, bare recitations of fact, and text like code where few word choices exist are all weaker cases for detection.
The inverse inference fails just as badly. A positive detection does not establish that someone handed the whole job to Claude. It is one input into thinking about a text’s provenance, not a determination of who wrote it or how.
Detection is also not something you can go run today. Anthropic has released a detection API in private preview, currently available to eligible organizations as required under EU law. Rollout to models released earlier is still in progress: the EU law includes a transition period for Anthropic models launched before August 2, 2026, and Anthropic says it is working to add watermarking to those as well. So do not assume every Claude output is uniformly watermarked right now. Anthropic’s help center is specific about it: as of September 2026 the models that carry the mark are Fable 5.1 and Mythos 5.1. If you drafted with an older model, there is no watermark on that text at all.
One thing that does change, and that is not the watermark: when Claude produces a supported file type — .png, .jpg, .svg — it attaches a C2PA content credential, a cryptographically signed note in the file’s metadata saying the file was made or processed with Claude. Anthropic is explicit that this is a different mechanism from watermarking: nothing inside the file changes, and the credential identifies no one. If you publish Claude-generated images, that metadata travels with them unless your image pipeline strips it.
EU Article 50: marking and disclosure are different obligations
This is where the watermark story connects to something publishers actually have to care about, and where the vocabulary gets mixed up.
Article 50(2) is about providers. Providers of AI systems generating synthetic audio, image, video, or text must ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. That is the obligation the watermark answers, and it falls on Anthropic, not on you.
Article 50(4) is about deployers. It covers deepfake disclosure, and separately requires deployers of a system generating or manipulating text published for the purpose of informing the public on matters of public interest to disclose that the text has been artificially generated or manipulated.
And then the part that matters most for a blog. Article 50(4) states that the obligation does not apply where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content.
These transparency obligations have applied since August 2, 2026 — this is current law, not an upcoming change. The European Commission published guidelines on the transparency obligations on July 20, 2026, and a Code of Practice on AI-generated content was finalized in June 2026.
I am not going to draw a legal conclusion here about any particular article, and neither should a blog post. Whether an obligation attaches depends on provider versus deployer, on the purpose of publication, and on facts specific to you. What I take from it is directional and, I think, uncontroversial: the exemption is built around editorial responsibility. Doing the editing is doing the compliance-relevant work.
SEO: there is no watermark penalty
The next thing I checked was whether watermarked text gets demoted in search.
As of September 2026, Google Search Central’s guidance on generative AI content says nothing about watermarks, SynthID, or any ranking signal derived from provenance. There is no documented watermark-based ranking factor. It does touch provenance in one narrow place — Merchant Center requires AI-generated product images to carry IPTC DigitalSourceType: TrainedAlgorithmicMedia metadata — but that is an ecommerce feed policy, not a Search ranking factor. Google’s position on production method is set out separately, in its 2023 post on AI-generated content: ranking systems reward content demonstrating experience, expertise, authoritativeness and trustworthiness, with the focus on “the quality of content, rather than how content is produced.”
What Google does say is that using generative AI to produce many pages without adding value for users may violate the spam policy on scaled content abuse — currently defined as generating many pages primarily to manipulate rankings rather than help users, regardless of how they were created. Google also suggests, as a recommendation rather than a requirement, that sharing information about how content was created can give readers useful context.
So neither “it has a watermark, therefore it is safe” nor “I removed the watermark, therefore it is safe” describes anything real. Both are the wrong axis.
I am not going back to edit old posts
Reading the announcement is not a reason to hunt through my archive for Claude-flavored sentences and rewrite them in bulk. I am also not going to push my articles through third-party detection services.
Detectors have limits, and a result cannot settle how something was produced. My published posts contain my own additions, edits, and fact-checking mixed in. Rewriting to satisfy a detector risks destroying the reasoning and readability that were there at the time — for no benefit.
What is worth doing is per-article: re-examine the evidence, the lived experience, and whether it is still current. When something is wrong or stale, fix it as ordinary editing.
What I will check once detection is generally available
When the detection API opens up beyond private preview, the things I want to know are: the official detection method, which models are covered, which languages, and how results vary with text length. Then I will run a small test on my own posts — short versus long, before editing versus after.
What I will not do is look for rewriting techniques that evade detection. The point is not concealing provenance. It is working out where in an AI-assisted process the human judgment belongs.
The five rules I am keeping
- Do not describe the watermark as hidden characters or as personally identifying
- Do not treat a detection result alone as establishing authorship or method
- Describe the technique as precisely as the official text does — no more cautious, no more certain
- Keep the provider’s marking obligation and the deployer’s disclosure obligation separate, and do not settle legal questions in a blog post
- Assume no watermark-specific search penalty exists; check accuracy, originality, and reader value instead
Summary
Claude’s text watermark embeds no hidden characters and no personal information. It is a statistical pattern in word selection, checkable with a key, described by Anthropic as a version of the SynthID-Text approach. Detection is weak on short text, weak on heavily edited text, and never sufficient on its own to establish who wrote something.
Nothing about that requires you to rewrite your archive. What it does reinforce is the thing EU Article 50 also singles out: separating what the AI produced from what you verified and what you actually experienced, and keeping a process where a human holds editorial responsibility before publication.
When the detection API becomes generally available I will test it against my own posts at that point’s specification. Until then, visible editorial responsibility beats worrying about an invisible mark.
References
- How Claude’s text watermarking works | Anthropic (August 14, 2026)
- How Claude marks AI-generated content | Anthropic Help Center
- Scalable watermarking for identifying large language model outputs | Nature
- Article 50: Transparency obligations for providers and deployers of certain AI systems | European Commission AI Act Service Desk
- Guidelines on transparency obligations for providers and deployers of AI systems | European Commission (July 20, 2026)
- Google Search’s guidance on using generative AI content on your website
- Spam policies for Google web search | Google Search Central