Anthropic has started embedding a cryptographic watermark into text generated by future Claude models, a change designed to help identify whether writing was produced by its AI. The company says the watermark is invisible to the human eye and doesn't alter Claude's normal output. The rollout is tied to the European Union's AI Act, which requires AI providers to mark generated text, and Anthropic signed the EU's Code of Practice on Transparency of AI-Generated Content in July 2026 alongside roughly 190 other signatories.
The company laid out the mechanics in a post on its website. Rather than picking the next word with a purely random number, the watermarked version of Claude bases that decision on a cryptographic key combined with the preceding text. That creates a subtle statistical pattern across a response that a human reader can't see but that anyone with the matching key can detect, allowing them to estimate the probability Claude produced the text.
No added cost or slowdown
Anthropic said the change carries no measurable downside. Internal testing found no difference in creativity, accuracy, or readability between watermarked and unwatermarked responses, according to the company, which also pointed to findings from Google DeepMind's original research on the underlying technique. A live-traffic test on a similar watermark showed no statistically significant shift in user satisfaction. The feature adds no extra tokens, meaning it doesn't slow Claude down or make it more expensive to use.
Where the watermark has limits
The watermark only works when a model is choosing among several equally valid options. Text with little room for variation - hard factual statements, precise code, math answers - carries a much weaker or nonexistent signal. Detection also grows less reliable on very short passages, since there's less pattern to analyze. Longer-form text gives a clearer read on whether Claude was involved.
Light edits to human writing may leave little to no detectable trace, since most of the original wording is untouched. A sufficiently heavy rewrite can remove the watermark entirely. At that point, Anthropic said, it becomes debatable whether the resulting text is still meaningfully AI-generated.
What the watermark can't tell you
The watermark can't be traced back to a specific user, account, or conversation, and it doesn't establish authorship or ownership over content. It only tells you the likelihood that Claude was involved in producing or editing the text. Anthropic distinguished the approach from third-party AI-detection tools, which typically rely on spotting stylistic patterns in AI writing rather than checking for an embedded signal tied to a private key.
Because Anthropic doesn't yet have a reliable way to apply the watermark only within the EU, the company is rolling it out globally and plans to extend it to older Claude models over the coming months. A separate API will soon let anyone check whether a piece of text carries Claude's watermark.
Why this matters for writers
If you use Claude to draft, edit, or brainstorm, clients and editors may soon be able to check your work against this watermark - not to identify you personally, but to estimate the chance that Claude touched the text. That's a meaningful shift for anyone who's been relying on the plausible deniability that comes with AI-generated prose. For now, the watermark is easiest to detect in longer, more varied writing, which happens to be the kind of work many writers produce every day. Knowing how the system works - and where it doesn't - can help you make deliberate choices about how much AI involvement you're comfortable with. If you're still getting comfortable with Claude's output, Claude AI Courses & Certifications can help you understand what the model does well and where its limits are. And for practical guidance on working with AI tools without losing your voice, AI for Writers covers the skills that matter as these detection systems spread.
Your membership also unlocks:
