Anthropic's EU watermark ships to every Claude user, everywhere

Anthropic used a blog post on Friday to explain the machine-readable watermarking it announced earlier in the week, built to comply with the European Union's AI Act ^[1]. The company says it is applying the marks to all of Claude's output at launch, globally rather than by region, because it does not have reliable ways to implement the change per jurisdiction ^[2].

That is the operationally interesting part. The AI Act requires providers to watermark AI-generated or manipulated audio, image, text and video output, applies to models released after August 2, and gives a grace period until December 2026 for previously released models ^[3]. Anthropic says the change affects output across all Claude products using models released after August 2, with watermarking added to older models over the coming months ^[4]. Its own framing is that all new models offered globally, not just in the EU, will mark generated content from day one ^[5].

The technique is an adaptation of Google DeepMind's SynthID-Text, described in a paper published in Nature ^[6]. Where a model would normally pick among several acceptable next words at random, a secret key steers the choice ^[7]. Anthropic's example: after "The weather today was cold and," either "grey" or "overcast" will do, and the key nudges the pick, while "Paris is the capital of" has one correct token and leaves no room to mark ^[8]. Anthropic says the difference between watermarked and unwatermarked text will not be distinguishable to readers, that it adds no hidden characters, and that generation is not slowed and costs no extra tokens ^[9][10]. Detection will be exposed as an API holding the keys, able to say whether Claude produced a given block of text ^[11].

For anyone shipping Claude output into a client deliverable, the edit path matters more than the generation path. Anthropic says it cannot distinguish text Claude wrote from text Claude edited, that translations are watermarked too, and that lightly editing watermarked text probably will not remove the mark; a full rewrite is the reliable way ^[12]. Short passages and light proofreading may fall below the detection threshold ^[13]. Code carries generally less watermarking because it usually requires exact output, and where there are no token choices the mark cannot be applied ^[14] - relevant given concerns that code may not be copyrightable without significant human input ^[15].

Ars Technica notes the EU guidance does not require marking where the system performs an assistive function for standard editing, such as grammar correction, or does not substantially alter the text or its meaning, and that a model-level watermark cannot tell wholesale generation from a comma fix ^[16]. How thorough the marking actually is will not be testable until the detection tool ships, which Anthropic says it plans to support eventually ^[17].

The churn signal is noisier than the headlines suggest. Gizmodo reports numerous X users claiming cancelled subscriptions, including the influencer John Ennis, who posted a cancellation screenshot on Saturday citing the "ridiculous watermark idea" ^[18]. Anthropic told Business Insider it has not seen an increase in cancellations since watermarking was announced ^[19], and some of the same posters cite an unrelated Wall Street Journal story about the CEO's wife's past business practices ^[20]. On Reddit, complaints ran alongside users defending the marks as a reasonable way to detect algorithmic output ^[21]. There is no verified churn number either way ^[1].

Watch for the detection API and its documented false-positive behaviour, whether pre-August models are retrofitted before the December 2026 deadline, and whether Anthropic ever narrows the global default once regional handling is feasible.