Anthropic Is Now Watermarking Claude's Text

If you write with Claude, your output now carries a signature you can’t see. Anthropic has started embedding an invisible watermark into the text its models generate, and it’s worth understanding exactly how that works before you assume it changes nothing about how you write.

How the Watermark Actually Works

According to Anthropic’s own announcement, the watermark doesn’t add hidden characters or alter your finished text after the fact. It changes something further upstream: the source of randomness Claude uses when choosing between multiple words that would work equally well in a sentence. Instead of picking randomly, Claude uses a cryptographic key together with the preceding words to guide those choices. The words it picks still look completely natural. But someone holding the right key can check whether a sequence of word choices matches the pattern Claude would produce, and use that as evidence the text came from Claude.

Because the watermark lives in the pattern of word choices themselves, not in extra characters, it travels with the text when it’s copied and pasted elsewhere. Anthropic says it can survive some editing too, though heavy rewriting will weaken or remove it, the same way editing a photo can strip out an image watermark.

Why Now: The EU AI Act

Regulation is driving this, not a voluntary industry trend. The EU AI Act’s Article 50 became enforceable on August 2, 2026, and it requires providers of generative AI systems to mark their outputs in a machine-readable way, so platforms, regulators, and users can eventually tell AI-generated content apart from human-written work. Anthropic joined roughly 190 signatories on the EU’s Code of Practice on Transparency of AI-Generated Content, and text watermarking is how it’s meeting that requirement. The company is rolling this out globally, not just for EU users, and older Claude models will get watermarking over the coming months under a transition period the law allows.

Anthropic isn’t acting alone here. It says its approach builds on SynthID-Text, the watermarking method Google DeepMind developed and later open-sourced. As we noted in our guide to Google’s SynthID rollout, Google’s own detection in Search only ever applied to content that a participating AI system chose to watermark. With a major model provider adopting the same underlying method, the pool of detectable AI text is about to grow.

Who Can Actually Check for the Watermark

This is the part that changes less than the headlines suggest. Detection isn’t public. Anthropic has launched a detection API, but it’s in private preview, limited to organizations the EU law specifically names as needing access: regulators, law enforcement, media outlets, fact-checkers, independent researchers, and educational and civil society groups. There’s no button available to the general public yet where you can paste in a paragraph and find out if Claude wrote it. Anthropic has said broader access is coming, but hasn’t committed to a date.

For now, the practical impact on everyday writers is smaller than it might seem. Your text carries a signature, but almost nobody can read it yet.

What This Means for Your Writing Workflow

None of this changes the core advice that’s mattered all along: using AI to draft doesn’t excuse you from making the final result good. A watermark tells someone with the right key that a machine picked certain words. It says nothing about whether the finished piece is accurate, original, or worth a reader’s time.

A few practical takeaways:

• Disclosure rules aren’t going away, and they’re not really about watermarks. If a platform or client requires you to disclose AI assistance, that’s an editorial or contractual obligation, not something a hidden signature enforces or replaces.
• Editing still matters, maybe more than ever. Since watermark strength can degrade with heavy rewriting, a Claude draft that goes through substantial human editing looks less like AI output by every measure that counts, not just the invisible one.
• Don’t treat “undetectable” as a strategy. Detection access is limited today. It’s expanding by design, driven by regulation with more signatories likely to follow. Building a content strategy around evading detection is a bet against where this is clearly headed.
• The real bar is still quality, not stealth. Whether or not a watermark can prove a machine helped write something, the content that performs, with readers and with search, is the content that’s accurate, well-researched, and genuinely useful. That was true before watermarking existed, and it stays true now.

Before you publish anything drafted with AI assistance, run it through a tool like GlyphFlow to check your word count, character limits, and Flesch Reading Ease score. A watermark can’t tell you if your writing is clear. Only editing does.