{"data":{"id":2190,"lang":"en","title":"Anthropic Adds Invisible Watermarks to All Claude-Generated Text","summary":"Anthropic now embeds invisible, statistically detectable watermarks in text from new Claude models to meet EU AI Act transparency rules.","category":"ai","category_label":"AI","published_at":"2026-08-15T02:01:34Z","updated_at":"2026-08-15T02:01:34Z","url":"https://techtools.cz/tools/news/article/2190-anthropic-adds-invisible-watermarks-to-all-claude-generated-text","api_url":"https://techtools.cz/news-api/v1/articles/2190?lang=en","markdown_url":"https://techtools.cz/news-api/v1/articles/2190.md?lang=en","image_url":"https://the-decoder.com/wp-content/uploads/2026/05/anthropic_logo_wall-1.png","word_count":208,"reading_time_minutes":2,"sources":[{"url":"https://the-decoder.com/anthropic-watermarks-all-claude-outputs-globally-with-marks-that-may-persist-through-some-editing/","domain":"the-decoder.com"},{"url":"https://www.forbes.com/sites/anishasircar/2026/08/13/claude-will-now-leave-a-watermark-on-everything-it-writes-what-does-that-mean/","domain":"forbes.com"}],"translation":{"lang":"cs","id":2263,"title":"Anthropic přidává neviditelné vodoznaky do veškerého textu generovaného Claude","url":"https://techtools.cz/tools/news/clanek/2263-anthropic-pridava-neviditelne-vodoznaky-do-veskereho-textu-generovaneho-claude","api_url":"https://techtools.cz/news-api/v1/articles/2263?lang=cs"},"content":"Anthropic has begun embedding invisible watermarks in text generated by new Claude models, starting with those released on or after August 2, 2026, including Claude Sonnet 4.6 and Claude Haiku 4.5. The rollout applies across every access channel, including Claude.ai, Claude Code and the API, and responds to transparency requirements in Article 50 of the EU AI Act, which Anthropic committed to under the bloc's AI Act Code of Practice.\n\nThe text watermark works by statistically biasing Claude's word choices according to a key held by Anthropic. No single word looks unusual, but the pattern becomes statistically detectable across enough generated text. Image and other file outputs instead carry signed provenance metadata following the C2PA standard. Anthropic says it plans to publish tools that let users and third parties check content for the marks, though it has not given a timeline.\n\nThe company is candid about the limits: a detected watermark does not prove Claude wrote a passage, since people often use it to edit or translate their own writing, and heavy editing, paraphrasing or very short outputs can erase the signal. Google DeepMind has taken a more open approach with its SynthID system, while OpenAI has reportedly held back a highly accurate detector for roughly two years."}}