# Claude will start watermarking generated text By Steven Van · 2026-08-14 The change, based on Google DeepMind's SynthID-Text, arrives to meet the EU AI Act's 2 August marking requirement. Future Claude models will add a text watermark, a change Anthropic and several other AI providers are making to comply with the EU AI Act, which as of 2 August requires AI providers serving the EU market to mark AI-generated content. The method, a version of Google DeepMind's SynthID-Text, works by biasing the random choice between equally valid words (like "overcast" versus "grey") using a hidden key rather than an arbitrary random number. Anthropic says this leaves no visible trace: nothing is added to the text, there are no hidden characters, and it doesn't require extra tokens or cost more. Anyone with the key can later estimate the likelihood that [Claude]() was involved in writing a passage. The watermark has limits Anthropic flagged directly. It can't confirm text is human-written, and it can't detect other AI models, even watermarked ones, since each uses a different key. Detection also gets weaker on short passages, on factual text with few word choices to work with, and on light proofreading edits, since the watermark can only attach to words Claude actually chose. It isn't applied to code where an exact output is required. Read the full explanation in [Anthropic's announcement](). [![]()]() ![Claude]() Claude AI assistant by Anthropic — write, analyze, code, and create with one of the most capable large language models available. [View Claude →]() [Original source]() [Read Claude will start watermarking generated text on Creators Toolbox]() --- Canonical source: https://creatorstoolbox.com/blog/claude-how-claude-s-text-watermark-works