Anthropic has confirmed that future Claude models will generate text carrying a watermark, a measure taken to comply with the EU AI Act.
The company says the watermarking method, based on Google DeepMind's SynthID-Text approach, has no impact on the quality, cost or speed of Claude's output and adds no hidden characters or extra tokens to the text. Instead, it works by altering how the model settles low-stakes word choices, such as between "overcast" and "grey", using a cryptographic key rather than arbitrary randomness. Anyone holding that key can then assess the probability that a piece of text was generated by Claude, though the watermark carries no information identifying a specific user, organisation or conversation.
Anthropic notes the approach has limitations. It cannot confirm whether text is human-written, cannot detect other AI systems' output, and is harder to detect in short passages or heavily factual and code-based content, where there are fewer equally valid word choices to draw on. Light human editing is unlikely to remove a watermark entirely, though a full rewrite would.
The change follows Anthropic's signing of the EU Code of Practice on Transparency of AI-Generated Content in July 2026, alongside roughly 190 other signatories including other major AI developers, who will implement their own watermarking systems. Watermarking is being applied globally rather than restricted to the EU, as Anthropic says it currently lacks a reliable way to scope the requirement by region. Older Claude models, launched before August 2, 2026, will receive watermarking over the coming months under a transition period built into the law.
